Published Lite
c718_fa_iq3kt_ryssig_q6k
Highlighted green throughout the tables.
Scope of this supplement
This page is intentionally large. The main production record explains the release decision. This supplement exists for readers who want the broad sweep: the score distributions, all candidate rows, recipe notes, and real-world harness rows.
Important boundary: this is every row in the summarized release datasets used here, not every unused generated recipe in the 882-entry candidate catalog. Untested generated recipes are not treated as results.
How the combinations started: the v1.1 mixed quants began as manual tensor-group recipes. We selected model regions such as full-attention Q/K/V/O, full-attention output, token/output tensors, RYS signal-path tensors, RYS FFN tensors, duplicate-boundary tensors, and late FFN control tensors, then paired those groups with ik-llama quant types. Scripts expanded, built, scheduled, and scored the recipe families, but the initial combinations were hand-designed hypotheses rather than an automatic per-tensor optimizer.
Published Plus
c154_fao_bf16_faqk_q8_lateffn_q8
Highlighted blue throughout the tables.
Original baseline
release_s010_iq4nl
Highlighted brown where present.
External baselines
unsloth_iq4nl, unsloth_q8_0
Highlighted gold where present.
What the scores mean
The score columns on this page come from two different evaluation layers. The hard-suite and comprehensive-suite rows are automated prompt-screen scores on a 0.0 to 1.0 scale. The real-world rows are verifier component counts such as 14/15 or 19/19.
Release interpretation: the prompt-suite scores were used for broad candidate discovery and narrowing. The real-world ClawCode scores carried more release weight because they tested whether a model could explore, edit, run tests, and repair inside an actual project workspace.
Prompt-suite score columns
| Column | Meaning | How to read it |
|---|---|---|
weighted or score | Primary automated prompt-suite score: 0.50 * habit_score + 0.35 * hard_score + 0.15 * mean_score - 0.05 * errors, clipped at zero. | Higher is better. It intentionally weights SignalLatch process habits more than raw puzzle accuracy. |
habit_score | Mean score for tasks in the signallatch_habits category. | Measures behavior retention: context gathering, command lifecycle control, repair loops, preserving user changes, calibrated uncertainty, and related agent habits. |
hard_score | Mean score for hard_reasoning, linked_reasoning, and hard_project tasks. | Measures exact reasoning, dependency-order reasoning, long-context instruction following, and practical planning under constraints. |
mean_score | Simple average across every task in that suite. | Useful as a sanity check when weighted score and category scores diverge. |
errors | Count of prompt requests that failed, timed out, or produced no usable response. | Each error lowers the weighted score and is treated as reliability evidence, not just a missing answer. |
size_gb | Candidate file size in decimal gigabytes. | Used to distinguish practical release files from larger variants that keep more tensors at high precision. |
score_minus_size_penalty | A screening-only size-adjusted score included in some summary tables. | Useful for finding practical candidates, but the final release decision used raw real-world evidence plus size/role tradeoffs rather than this column alone. |
score_per_gb | Weighted score divided by candidate size. | Efficiency hint. It is not a quality score by itself. |
How individual prompt tasks were scored
Exact tasks used answer-pattern checks plus required-term checks. The exact-task score was 0.78 answer-pattern coverage plus 0.22 required-term coverage, with penalties for forbidden terms and excessive repeated 5-grams.
Rubric tasks used category keywords. Each rubric category passed if the answer contained at least one term from that category. The task score was the mean of category passes, with penalties for very short answers, forbidden terms, and excessive repeated 5-grams.
Because this is an automated screen, a high prompt-suite score means "passed the public scoring signals for this suite"; it does not mean a human would prefer that model in every coding session. That is why the project-level ClawCode harnesses were added after the broad screen.
Stable gate vs comprehensive gate
| Suite | Rows shown here | Prompt tasks | Purpose |
|---|---|---|---|
| Stable promotion screen | 163 | 12 | Fast gate used as the stable public count: 163 * 12 = 1,956 prompt-level scored results. |
| Comprehensive summary | 440 | 16 | Broader later sweep with the same 12 core tasks plus four long-context retention tasks. |
| Real-world ClawCode harnesses | 32 | Project verifiers | Finalist comparison using actual workspaces and component checks. |
Prompt-suite task catalog
| Task ID | Category | Scoring type | What it tested |
|---|---|---|---|
crt_exact_1136 | hard_reasoning | Exact | Chinese-remainder arithmetic, required modular checks, and a final exact answer of 1136. |
state_machine_exact_8_minus5 | hard_reasoning | Exact | Six-step conditional state tracking, avoiding skipped updates, ending with x=8 and y=-5. |
binary_strings_exact_126 | hard_reasoning | Exact | Combinatorics with no-adjacent-ones gap reasoning, ending with 126. |
dependency_chain_project_order | linked_reasoning | Rubric | Evidence-first pipeline debugging, dependency order, validation, and avoiding the wrong parser-rewrite scope. |
deadlock_ordering_plan | hard_reasoning | Rubric | Cycle diagnosis, lock-ordering rule, regression tests, and validation for a deadlock-style failure. |
agent_context_sufficiency | signallatch_habits | Rubric | Goal tracking, context gaps, evidence to inspect, and a stop condition before editing a vague repo issue. |
command_lifecycle_control | signallatch_habits | Rubric | Starting, monitoring, polling, logging, stopping, and cleaning up a long GPU/model evaluation without killing unrelated services. |
repair_from_failure_signal | signallatch_habits | Rubric | Using a concrete focused-test failure to plan minimal repair and choose when to rerun broad tests. |
restrained_delegation | signallatch_habits | Rubric | Keeping blocking stack-trace work local while delegating only independent, non-overlapping side work. |
preserve_user_changes | signallatch_habits | Rubric | Dirty-worktree discipline: do not revert unrelated user changes, ignore unrelated files, and work with same-file changes when needed. |
complex_project_plan_hard | hard_project | Rubric | Planning an automated quant sweep under two-server, GPU-pool, cleanup, and tiered-test constraints. |
uncertainty_and_evidence | signallatch_habits | Rubric | Calibrated model-selection answer when style evidence and exact-reasoning evidence disagree. |
long_exact_redwood_12289 | hard_reasoning | Exact | Long-context active-control-packet retention with noisy irrelevant context, ending with 12289. |
long_linked_topology_order | linked_reasoning | Exact | Long-context task-order retention, requiring READ > STOP > PATCH > BUILD > EVAL > PRUNE. |
long_habit_preserve_changes | signallatch_habits | Rubric | Long-context retention of inspect-before-editing, no destructive git operations, user-change preservation, conflict scope, and targeted validation. |
long_project_gate_7046 | hard_project | Exact | Long project-transcript retention of goal tracking, evidence-before-editing, command lifecycle control, scoped validation, and final exact answer 7046. |
Real-world harness scores
A real-world score such as 14/15 means the verifier found 14 passing components out of 15. Components included visible tests, hidden tests, and structural checks such as keeping scope focused, preserving tests, avoiding dependency churn, preserving old retrieval paths, or adding deterministic verification paths.
The ClawCode process return code was recorded but was not the quality score. The verifier inspected the final workspace and counted behavior-specific components, because a run can produce a useful patch while still ending with a non-zero tool return or a missed hidden edge case.
Massive summary graphs
The bar lists below are full result lists, not top-N excerpts. They are scrollable because the comprehensive run has 440 rows.
All comprehensive candidates by weighted score
All stable promotion-screen candidates by weighted score
Every real-world harness result
These are the project-level ClawCode harness rows. They are the closest tests to the natural use case: work inside a project, infer the files, implement, test, and repair.
| Candidate | KritaLite | SyncBoard | Legacy-codebase migration | Recipe spec |
|---|---|---|---|---|
c718_fa_iq3kt_ryssig_q6k |
14/15 93.3% |
16/19 84.2% |
19/19 100.0% |
fa=iq3_kt, rys_signal=q6_K |
unsloth_q8_0 |
14/15 93.3% |
19/19 100.0% |
18/19 94.7% |
|
c154_fao_bf16_faqk_q8_lateffn_q8 |
13/15 86.7% |
18/19 94.7% |
19/19 100.0% |
fa_o=bf16, fa_qk=q8_0, late_ffn=q8_0 |
c204_fao_iq4kss |
13/15 86.7% |
19/19 100.0% |
8/19 42.1% |
fa_o=iq4_kss |
c311_outtok_iq3kt |
12/15 80.0% |
not run | 19/19 100.0% |
output_token=iq3_kt |
c373_fa_iq3kr4 |
13/15 86.7% |
18/19 94.7% |
18/19 94.7% |
fa=iq3_k_r4 |
c281_lateffn_iq3kt |
12/15 80.0% |
not run | 18/19 94.7% |
late_ffn=iq3_kt |
c404_faqkv_iq5kr4 |
not run | not run | 18/19 94.7% |
fa_qkv=iq5_k_r4 |
c76_fa_q6k_outtok_bf16 |
14/15 93.3% |
18/19 94.7% |
8/19 42.1% |
fa=q6_K, output_token=bf16 |
unsloth_iq4nl |
14/15 93.3% |
18/19 94.7% |
17/19 89.5% |
|
release_s010_iq4nl |
12/15 80.0% |
18/19 94.7% |
16/19 84.2% |
|
c317_out_iq5k |
not run | not run | 17/19 89.5% |
output_only=iq5_k |
c05_all_fullattn_bf16 |
13/15 86.7% |
not run | not run | ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 |
c431_ryspath_q40r8 |
not run | not run | 14/19 73.7% |
rys_path=q4_0_r8 |
| Suite | Candidate | Components | Score | Elapsed s | Pool | Context | Claw rc | Verify rc | Files | Tests |
|---|---|---|---|---|---|---|---|---|---|---|
| KritaLite | c718_fa_iq3kt_ryssig_q6k | 14/15 | 93.3% | 448 | high | 160000 | 0 | 1 | 0 | |
| KritaLite | c76_fa_q6k_outtok_bf16 | 14/15 | 93.3% | 1195 | small_b | 160000 | 0 | 1 | 0 | |
| KritaLite | unsloth_iq4nl | 14/15 | 93.3% | 246 | high | 160000 | 0 | 1 | 0 | |
| KritaLite | unsloth_q8_0 | 14/15 | 93.3% | 1495 | small_a | 160000 | 0 | 1 | 0 | |
| KritaLite | c05_all_fullattn_bf16 | 13/15 | 86.7% | 3187 | small_a | 160000 | 0 | 1 | 0 | |
| KritaLite | c154_fao_bf16_faqk_q8_lateffn_q8 | 13/15 | 86.7% | 779 | high | 160000 | 0 | 1 | 0 | |
| KritaLite | c204_fao_iq4kss | 13/15 | 86.7% | 1924 | small_a | 160000 | 0 | 1 | 0 | |
| KritaLite | c373_fa_iq3kr4 | 13/15 | 86.7% | 837 | small_a | 160000 | 0 | 1 | 0 | |
| KritaLite | c281_lateffn_iq3kt | 12/15 | 80.0% | 2130 | small_b | 160000 | 0 | 1 | 0 | |
| KritaLite | c311_outtok_iq3kt | 12/15 | 80.0% | 1933 | small_b | 160000 | 0 | 1 | 0 | |
| KritaLite | release_s010_iq4nl | 12/15 | 80.0% | 568 | high | 160000 | 0 | 1 | 0 | |
| Legacy-codebase migration | c154_fao_bf16_faqk_q8_lateffn_q8 | 19/19 | 100.0% | 7041 | high | 160000 | 1 | 0 | 5 | 5/5 |
| Legacy-codebase migration | c311_outtok_iq3kt | 19/19 | 100.0% | 10906 | small_a | 160000 | 0 | 0 | 8 | 6/6 |
| Legacy-codebase migration | c718_fa_iq3kt_ryssig_q6k | 19/19 | 100.0% | 2100 | high | 160000 | 1 | 0 | 8 | 6/6 |
| Legacy-codebase migration | c281_lateffn_iq3kt | 18/19 | 94.7% | 1497 | small_a | 160000 | 0 | 0 | 4 | 3/3 |
| Legacy-codebase migration | c373_fa_iq3kr4 | 18/19 | 94.7% | 7742 | small_b | 160000 | 0 | 0 | 5 | 5/5 |
| Legacy-codebase migration | c404_faqkv_iq5kr4 | 18/19 | 94.7% | 2045 | small_a | 160000 | 1 | 0 | 5 | 3/4 |
| Legacy-codebase migration | unsloth_q8_0 | 18/19 | 94.7% | 1816 | high | 160000 | 1 | 0 | 5 | 3/3 |
| Legacy-codebase migration | c317_out_iq5k | 17/19 | 89.5% | 1841 | small_b | 160000 | 0 | 0 | 3 | 3/3 |
| Legacy-codebase migration | unsloth_iq4nl | 17/19 | 89.5% | 2062 | high | 160000 | 1 | 0 | 5 | 5/5 |
| Legacy-codebase migration | release_s010_iq4nl | 16/19 | 84.2% | 1611 | high | 160000 | 1 | 0 | 5 | 3/3 |
| Legacy-codebase migration | c431_ryspath_q40r8 | 14/19 | 73.7% | 14400 | small_b | 160000 | 124 | 1 | 1 | 3/3 |
| Legacy-codebase migration | c204_fao_iq4kss | 8/19 | 42.1% | 4397 | small_a | 160000 | 0 | 1 | 1 | 3/3 |
| Legacy-codebase migration | c76_fa_q6k_outtok_bf16 | 8/19 | 42.1% | 1466 | high | 160000 | 1 | 1 | 1 | 3/3 |
| SyncBoard | c204_fao_iq4kss | 19/19 | 100.0% | 1240 | small_b | 160000 | 0 | 0 | 0 | |
| SyncBoard | unsloth_q8_0 | 19/19 | 100.0% | 647 | high | 160000 | 0 | 0 | 0 | |
| SyncBoard | c154_fao_bf16_faqk_q8_lateffn_q8 | 18/19 | 94.7% | 3724 | small_a | 160000 | 1 | 1 | 0 | |
| SyncBoard | c373_fa_iq3kr4 | 18/19 | 94.7% | 348 | high | 160000 | 0 | 1 | 0 | |
| SyncBoard | c76_fa_q6k_outtok_bf16 | 18/19 | 94.7% | 601 | small_a | 160000 | 0 | 1 | 0 | |
| SyncBoard | release_s010_iq4nl | 18/19 | 94.7% | 804 | small_b | 160000 | 0 | 1 | 0 | |
| SyncBoard | unsloth_iq4nl | 18/19 | 94.7% | 1145 | small_a | 160000 | 0 | 1 | 0 | |
| SyncBoard | c718_fa_iq3kt_ryssig_q6k | 16/19 | 84.2% | 2350 | high | 160000 | 0 | 1 | 0 |
Comprehensive candidate table
This table contains all 440 rows in summary_signallatch_comprehensive_eval_latest.json, joined with recipe descriptions from the candidate catalog when available.
| Rank | Candidate | Weighted | Habit | Hard | Mean | Size | Errors | Recipe spec | Description / exact rules |
|---|---|---|---|---|---|---|---|---|---|
| 1 | c373_fa_iq3kr4 |
0.8935 | 0.8786 | 0.9130 | 0.8979 | 16.29 GB | 0 | fa=iq3_k_r4 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq3_k_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq3_k_r4 |
| 2 | c281_lateffn_iq3kt |
0.8916 | 0.8786 | 0.9085 | 0.8954 | 16.43 GB | 0 | late_ffn=iq3_kt | Mixed manual probe: late FFN gate/down control tensors at iq3_kt.rules^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq3_kt |
| 3 | c718_fa_iq3kt_ryssig_q6k |
0.8822 | 0.9286 | 0.8219 | 0.8685 | 16.40 GB | 0 | fa=iq3_kt, rys_signal=q6_K | Mixed manual probe: all full-attention Q/K/V/O tensors at iq3_kt; RYS 15..24 linear-attention and SSM signal path at q6_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq3_kt ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q6_K |
| 4 | c204_fao_iq4kss |
0.8761 | 0.9143 | 0.8263 | 0.8648 | 16.52 GB | 0 | fa_o=iq4_kss | Mixed manual probe: full-attention output tensors at iq4_kss.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=iq4_kss |
| 5 | c311_outtok_iq3kt |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 15.79 GB | 0 | output_token=iq3_kt | Mixed manual probe: token embedding and output tensors at iq3_kt.rules^output\.weight=iq3_kt ^token_embd\.weight=iq3_kt |
| 6 | c317_out_iq5k |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.39 GB | 0 | output_only=iq5_k | Mixed manual probe: output tensor at iq5_k.rules^output\.weight=iq5_k |
| 7 | c254_rysffn_iq4kss |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.39 GB | 0 | rys_ffn=iq4_kss | Mixed manual probe: RYS 15..24 FFN tensors at iq4_kss.rules^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=iq4_kss |
| 8 | c500_fa_iq4ks_ryssig_iq4kss |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.44 GB | 0 | fa=iq4_ks, rys_signal=iq4_kss | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_ks; RYS 15..24 linear-attention and SSM signal path at iq4_kss.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_ks ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_kss |
| 9 | c234_bndall_iq4kss |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.46 GB | 0 | boundary_all=iq4_kss | Mixed manual probe: RYS entry, exit, and duplicate-boundary attention/SSM/FFN tensors at iq4_kss.rules^blk\.(15|19|20|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=iq4_kss |
| 10 | c253_rysffn_iq4ks |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.47 GB | 0 | rys_ffn=iq4_ks | Mixed manual probe: RYS 15..24 FFN tensors at iq4_ks.rules^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=iq4_ks |
| 11 | c169_fa_iq4xs |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.48 GB | 0 | fa=iq4_xs | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_xs.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_xs |
| 12 | c274_lateffn_iq4kss |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.51 GB | 0 | late_ffn=iq4_kss | Mixed manual probe: late FFN gate/down control tensors at iq4_kss.rules^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq4_kss |
| 13 | c456_latectl_iq4xsr8 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.51 GB | 0 | late_control=iq4_xs_r8 | Mixed manual probe: late attention plus FFN gate/down control tensors at iq4_xs_r8.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_xs_r8 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq4_xs_r8 |
| 14 | c213_ryspath_iq4ks |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.52 GB | 0 | rys_path=iq4_ks | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at iq4_ks.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_ks |
| 15 | c364_duppath_iq4kt |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.52 GB | 0 | dup_path=iq4_kt | Mixed manual probe: duplicated-side RYS 20..24 attention, linear-attention, and SSM path at iq4_kt.rules^blk\.(20|21|22|23|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_kt |
| 16 | c201_fav_iq3kt |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.53 GB | 0 | fa_v=iq3_kt | Mixed manual probe: full-attention V tensors at iq3_kt.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=iq3_kt |
| 17 | c273_lateffn_iq4ks |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.53 GB | 0 | late_ffn=iq4_ks | Mixed manual probe: late FFN gate/down control tensors at iq4_ks.rules^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq4_ks |
| 18 | c200_fav_iq4kt |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.54 GB | 0 | fa_v=iq4_kt | Mixed manual probe: full-attention V tensors at iq4_kt.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=iq4_kt |
| 19 | c339_earlyattn_iq4xs |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.54 GB | 0 | early_attn=iq4_xs | Mixed manual probe: early full-attention routing tensors at iq4_xs.rules^blk\.(3|7|11|15)\.attn_(q|k|v|output)\.weight=iq4_xs |
| 20 | c193_fav_iq4ks |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.54 GB | 0 | fa_v=iq4_ks | Mixed manual probe: full-attention V tensors at iq4_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=iq4_ks |
| 21 | c369_fa_iq4kr4 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.54 GB | 0 | fa=iq4_k_r4 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_k_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_k_r4 |
| 22 | c336_earlyattn_iq4k |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.55 GB | 0 | early_attn=iq4_k | Mixed manual probe: early full-attention routing tensors at iq4_k.rules^blk\.(3|7|11|15)\.attn_(q|k|v|output)\.weight=iq4_k |
| 23 | c423_ryspath_iq4nlr4 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.55 GB | 0 | rys_path=iq4_nl_r4 | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at iq4_nl_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_nl_r4 |
| 24 | c431_ryspath_q40r8 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.55 GB | 0 | rys_path=q4_0_r8 | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at q4_0_r8.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q4_0_r8 |
| 25 | c206_fao_iq4k |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.55 GB | 0 | fa_o=iq4_k | Mixed manual probe: full-attention output tensors at iq4_k.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=iq4_k |
| 26 | c226_ryssig_iq4k |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.55 GB | 0 | rys_signal=iq4_k | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at iq4_k.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_k |
| 27 | c397_faqk_q40r8 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.55 GB | 0 | fa_qk=q4_0_r8 | Mixed manual probe: full-attention Q/K tensors at q4_0_r8.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=q4_0_r8 |
| 28 | c205_fao_iq5ks |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.61 GB | 0 | fa_o=iq5_ks | Mixed manual probe: full-attention output tensors at iq5_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=iq5_ks |
| 29 | c433_ryspath_q5kr4 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.69 GB | 0 | rys_path=q5_k_r4 | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at q5_k_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q5_k_r4 |
| 30 | c649_fa_iq5k_rysffn_iq4ks |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.70 GB | 0 | fa=iq5_k, rys_ffn=iq4_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_k; RYS 15..24 FFN tensors at iq4_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_k ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=iq4_ks |
| 31 | c21_fao_q6k |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.70 GB | 0 | fa_o=q6_K | Mixed manual probe: full-attention output tensors at q6_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=q6_K |
| 32 | c208_fao_iq6k |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.71 GB | 0 | fa_o=iq6_k | Mixed manual probe: full-attention output tensors at iq6_k.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=iq6_k |
| 33 | c404_faqkv_iq5kr4 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.71 GB | 0 | fa_qkv=iq5_k_r4 | Mixed manual probe: full-attention Q/K/V tensors at iq5_k_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=iq5_k_r4 |
| 34 | c610_fa_iq4k_ryssig_q6k |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.72 GB | 0 | fa=iq4_k, rys_signal=q6_K | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_k; RYS 15..24 linear-attention and SSM signal path at q6_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_k ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q6_K |
| 35 | c655_fa_iq5k_bndffn_iq4ks |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.75 GB | 0 | fa=iq5_k, boundary_ffn=iq4_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_k; RYS entry, exit, and duplicate-boundary FFN tensors at iq4_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_k ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=iq4_ks |
| 36 | c167_fa_iq5k |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.78 GB | 0 | fa=iq5_k | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_k.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_k |
| 37 | c370_fa_iq5kr4 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.78 GB | 0 | fa=iq5_k_r4 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_k_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_k_r4 |
| 38 | c379_fa_q50r4 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.78 GB | 0 | fa=q5_0_r4 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_0_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_0_r4 |
| 39 | c22_fao_q8 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 16.84 GB | 0 | fa_o=q8_0 | Mixed manual probe: full-attention output tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=q8_0 |
| 40 | c391_faqk_q8r8 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 17.17 GB | 0 | fa_qk=q8_0_r8 | Mixed manual probe: full-attention Q/K tensors at q8_0_r8.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=q8_0_r8 |
| 41 | c616_fa_iq4k_rysffn_q6k |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 17.23 GB | 0 | fa=iq4_k, rys_ffn=q6_K | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_k; RYS 15..24 FFN tensors at q6_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_k ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q6_K |
| 42 | c587_fa_iq5ks_bndffn_q8 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 17.26 GB | 0 | fa=iq5_ks, boundary_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_ks; RYS entry, exit, and duplicate-boundary FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 |
| 43 | c77_fa_q6k_ryssig_q8 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 17.38 GB | 0 | fa=q6_K, rys_signal=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K; RYS 15..24 linear-attention and SSM signal path at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 44 | c666_fa_iq5k_lateffn_bf16 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 17.80 GB | 0 | fa=iq5_k, late_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_k; late FFN gate/down control tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_k ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=bf16 |
| 45 | c93_fa_q8_bndffn_q8 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 18.02 GB | 0 | fa=q8_0, boundary_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; RYS entry, exit, and duplicate-boundary FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 |
| 46 | c154_fao_bf16_faqk_q8_lateffn_q8 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 18.34 GB | 0 | fa_o=bf16, fa_qk=q8_0, late_ffn=q8_0 | Mixed manual probe: full-attention output tensors at bf16; full-attention Q/K tensors at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=bf16 ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 47 | c06_output_token_q8_fullattn_q8 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 18.43 GB | 0 | ^output\.weight=q8_0; ^token_embd\.weight=q8_0; ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 | Protect expression/input surfaces plus full-attention routing: token/output Q8_0 and full-attention Q/K/V/O Q8_0.rules^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 |
| 48 | c05_all_fullattn_bf16 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 19.26 GB | 0 | ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 | Full-attention ceiling probe: all full-attention Q/K/V/O tensors remain BF16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 |
| 49 | c56_outtok_bf16 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 19.88 GB | 0 | output_token=bf16 | Mixed manual probe: token embedding and output tensors at bf16.rules^output\.weight=bf16 ^token_embd\.weight=bf16 |
| 50 | c145_fa_bf16_outtok_q8_ryssig_q8 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 20.55 GB | 0 | fa=bf16, output_token=q8_0, rys_signal=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; token embedding and output tensors at q8_0; RYS 15..24 linear-attention and SSM signal path at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 51 | c147_fa_bf16_outtok_q8_bndffn_q8 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 20.73 GB | 0 | fa=bf16, output_token=q8_0, boundary_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; token embedding and output tensors at q8_0; RYS entry, exit, and duplicate-boundary FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 |
| 52 | c106_fa_bf16_bndffn_bf16 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 20.79 GB | 0 | fa=bf16, boundary_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; RYS entry, exit, and duplicate-boundary FFN tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=bf16 |
| 53 | c88_fa_q8_outtok_bf16 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 20.81 GB | 0 | fa=q8_0, output_token=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; token embedding and output tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^output\.weight=bf16 ^token_embd\.weight=bf16 |
| 54 | c100_fa_bf16_outtok_bf16 |
0.8741 | 0.9143 | 0.8219 | 0.8623 | 22.58 GB | 0 | fa=bf16, output_token=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; token embedding and output tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^output\.weight=bf16 ^token_embd\.weight=bf16 |
| 55 | c172_fa_mxfp4 |
0.8734 | 0.9143 | 0.8202 | 0.8614 | 16.48 GB | 0 | fa=mxfp4 | Mixed manual probe: all full-attention Q/K/V/O tensors at mxfp4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=mxfp4 |
| 56 | c192_faqkv_mxfp4 |
0.8734 | 0.9143 | 0.8202 | 0.8614 | 16.50 GB | 0 | fa_qkv=mxfp4 | Mixed manual probe: full-attention Q/K/V tensors at mxfp4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=mxfp4 |
| 57 | c239_bndall_iq4xs |
0.8722 | 0.9143 | 0.8174 | 0.8598 | 16.51 GB | 0 | boundary_all=iq4_xs | Mixed manual probe: RYS entry, exit, and duplicate-boundary attention/SSM/FFN tensors at iq4_xs.rules^blk\.(15|19|20|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=iq4_xs |
| 58 | c377_fa_q60 |
0.8722 | 0.9143 | 0.8174 | 0.8598 | 17.01 GB | 0 | fa=q6_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_0 |
| 59 | c413_faqkv_q50r4 |
0.8715 | 0.9143 | 0.8157 | 0.8589 | 16.71 GB | 0 | fa_qkv=q5_0_r4 | Mixed manual probe: full-attention Q/K/V tensors at q5_0_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=q5_0_r4 |
| 60 | c207_fao_iq5k |
0.8683 | 0.9143 | 0.8085 | 0.8548 | 16.63 GB | 0 | fa_o=iq5_k | Mixed manual probe: full-attention output tensors at iq5_k.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=iq5_k |
| 61 | c312_outtok_mxfp4 |
0.8678 | 0.8786 | 0.8537 | 0.8646 | 16.15 GB | 0 | output_token=mxfp4 | Mixed manual probe: token embedding and output tensors at mxfp4.rules^output\.weight=mxfp4 ^token_embd\.weight=mxfp4 |
| 62 | c300_finalffn_iq4kt |
0.8664 | 0.9143 | 0.8041 | 0.8523 | 16.51 GB | 0 | final_ffn=iq4_kt | Mixed manual probe: final FFN tensors at iq4_kt.rules^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=iq4_kt |
| 63 | c474_outtok_iq4nlr4 |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.23 GB | 0 | output_token=iq4_nl_r4 | Mixed manual probe: token embedding and output tensors at iq4_nl_r4.rules^output\.weight=iq4_nl_r4 ^token_embd\.weight=iq4_nl_r4 |
| 64 | c691_fa_iq4kt_bndffn_iq4ks |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.39 GB | 0 | fa=iq4_kt, boundary_ffn=iq4_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kt; RYS entry, exit, and duplicate-boundary FFN tensors at iq4_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kt ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=iq4_ks |
| 65 | c184_faqkv_iq4kss |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.46 GB | 0 | fa_qkv=iq4_kss | Mixed manual probe: full-attention Q/K/V tensors at iq4_kss.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=iq4_kss |
| 66 | c330_tok_iq4kt |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.48 GB | 0 | token_only=iq4_kt | Mixed manual probe: token embedding tensor at iq4_kt.rules^token_embd\.weight=iq4_kt |
| 67 | c174_faqk_iq4kss |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.48 GB | 0 | fa_qk=iq4_kss | Mixed manual probe: full-attention Q/K tensors at iq4_kss.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=iq4_kss |
| 68 | c244_bndffn_iq4kss |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.49 GB | 0 | boundary_ffn=iq4_kss | Mixed manual probe: RYS entry, exit, and duplicate-boundary FFN tensors at iq4_kss.rules^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=iq4_kss |
| 69 | c329_tok_iq4xs |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.52 GB | 0 | token_only=iq4_xs | Mixed manual probe: token embedding tensor at iq4_xs.rules^token_embd\.weight=iq4_xs |
| 70 | c219_ryspath_iq4xs |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.52 GB | 0 | rys_path=iq4_xs | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at iq4_xs.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_xs |
| 71 | c210_fao_iq4kt |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.52 GB | 0 | fa_o=iq4_kt | Mixed manual probe: full-attention output tensors at iq4_kt.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=iq4_kt |
| 72 | c252_bndffn_mxfp4 |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.52 GB | 0 | boundary_ffn=mxfp4 | Mixed manual probe: RYS entry, exit, and duplicate-boundary FFN tensors at mxfp4.rules^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=mxfp4 |
| 73 | c279_lateffn_iq4xs |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.53 GB | 0 | late_ffn=iq4_xs | Mixed manual probe: late FFN gate/down control tensors at iq4_xs.rules^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq4_xs |
| 74 | c209_fao_iq4xs |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.54 GB | 0 | fa_o=iq4_xs | Mixed manual probe: full-attention output tensors at iq4_xs.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=iq4_xs |
| 75 | c203_fao_iq4ks |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.54 GB | 0 | fa_o=iq4_ks | Mixed manual probe: full-attention output tensors at iq4_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=iq4_ks |
| 76 | c166_fa_iq4k |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.54 GB | 0 | fa=iq4_k | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_k.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_k |
| 77 | c472_outtok_iq5kr4 |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.54 GB | 0 | output_token=iq5_k_r4 | Mixed manual probe: token embedding and output tensors at iq5_k_r4.rules^output\.weight=iq5_k_r4 ^token_embd\.weight=iq5_k_r4 |
| 78 | c267_lateattn_iq5k |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.62 GB | 0 | late_attn=iq5_k | Mixed manual probe: late full-attention control tensors at iq5_k.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_k |
| 79 | c583_fa_iq5ks_bndffn_iq4ks |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.69 GB | 0 | fa=iq5_ks, boundary_ffn=iq4_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_ks; RYS entry, exit, and duplicate-boundary FFN tensors at iq4_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=iq4_ks |
| 80 | c639_fa_iq5k_outtok_iq5ks |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.69 GB | 0 | fa=iq5_k, output_token=iq5_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_k; token embedding and output tensors at iq5_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_k ^output\.weight=iq5_ks ^token_embd\.weight=iq5_ks |
| 81 | c464_latectl_q50r4 |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.71 GB | 0 | late_control=q5_0_r4 | Mixed manual probe: late attention plus FFN gate/down control tensors at q5_0_r4.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_0_r4 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q5_0_r4 |
| 82 | c467_latectl_q5kr4 |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.71 GB | 0 | late_control=q5_k_r4 | Mixed manual probe: late attention plus FFN gate/down control tensors at q5_k_r4.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_k_r4 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q5_k_r4 |
| 83 | c382_fa_q5kr4 |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.78 GB | 0 | fa=q5_k_r4 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_k_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_k_r4 |
| 84 | c491_ryssig_q8kv |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.86 GB | 0 | rys_signal=q8_KV | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at q8_KV.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_KV |
| 85 | c604_fa_iq4k_outtok_q6k |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.87 GB | 0 | fa=iq4_k, output_token=q6_K | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_k; token embedding and output tensors at q6_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_k ^output\.weight=q6_K ^token_embd\.weight=q6_K |
| 86 | c178_faqk_iq6k |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.88 GB | 0 | fa_qk=iq6_k | Mixed manual probe: full-attention Q/K tensors at iq6_k.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=iq6_k |
| 87 | c111_fav_bf16_fao_q8 |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 16.96 GB | 0 | fa_v=bf16, fa_o=q8_0 | Mixed manual probe: full-attention V tensors at bf16; full-attention output tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=bf16 ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=q8_0 |
| 88 | c544_fa_iq4kss_rysffn_q6k |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 17.12 GB | 0 | fa=iq4_kss, rys_ffn=q6_K | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kss; RYS 15..24 FFN tensors at q6_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kss ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q6_K |
| 89 | c23_fao_bf16 |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 17.37 GB | 0 | fa_o=bf16 | Mixed manual probe: full-attention output tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=bf16 |
| 90 | c564_fa_iq4kss_finalffn_bf16 |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 17.58 GB | 0 | fa=iq4_kss, final_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kss; final FFN tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kss ^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=bf16 |
| 91 | c569_fa_iq5ks_outtok_q8 |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 17.67 GB | 0 | fa=iq5_ks, output_token=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_ks; token embedding and output tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks ^output\.weight=q8_0 ^token_embd\.weight=q8_0 |
| 92 | c152_fav_bf16_fao_bf16_bndffn_q8 |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 18.03 GB | 0 | fa_v=bf16, fa_o=bf16, boundary_ffn=q8_0 | Mixed manual probe: full-attention V tensors at bf16; full-attention output tensors at bf16; RYS entry, exit, and duplicate-boundary FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=bf16 ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=bf16 ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 |
| 93 | c151_fav_bf16_fao_bf16_outtok_q8 |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 18.44 GB | 0 | fa_v=bf16, fa_o=bf16, output_token=q8_0 | Mixed manual probe: full-attention V tensors at bf16; full-attention output tensors at bf16; token embedding and output tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=bf16 ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=bf16 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 |
| 94 | c141_fa_q8_outtok_q8_bndffn_q8 |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 18.97 GB | 0 | fa=q8_0, output_token=q8_0, boundary_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; token embedding and output tensors at q8_0; RYS entry, exit, and duplicate-boundary FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 |
| 95 | c64_fa_q5k_outtok_bf16 |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 20.11 GB | 0 | fa=q5_K, output_token=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K; token embedding and output tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K ^output\.weight=bf16 ^token_embd\.weight=bf16 |
| 96 | c76_fa_q6k_outtok_bf16 |
0.8645 | 0.9143 | 0.7996 | 0.8498 | 20.36 GB | 0 | fa=q6_K, output_token=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K; token embedding and output tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^output\.weight=bf16 ^token_embd\.weight=bf16 |
| 97 | c751_fa_mxfp4_ryssig_iq4ks |
0.8638 | 0.9143 | 0.7980 | 0.8489 | 16.46 GB | 0 | fa=mxfp4, rys_signal=iq4_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at mxfp4; RYS 15..24 linear-attention and SSM signal path at iq4_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=mxfp4 ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_ks |
| 98 | c592_fa_iq5ks_lateffn_q6k |
0.8635 | 0.9143 | 0.7974 | 0.8485 | 16.91 GB | 0 | fa=iq5_ks, late_ffn=q6_K | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_ks; late FFN gate/down control tensors at q6_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q6_K |
| 99 | c322_out_mxfp4 |
0.8581 | 0.8786 | 0.8315 | 0.8521 | 16.19 GB | 0 | output_only=mxfp4 | Mixed manual probe: output tensor at mxfp4.rules^output\.weight=mxfp4 |
| 100 | c407_faqkv_iq3kr4 |
0.8575 | 0.8786 | 0.8300 | 0.8512 | 16.37 GB | 0 | fa_qkv=iq3_k_r4 | Mixed manual probe: full-attention Q/K/V tensors at iq3_k_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=iq3_k_r4 |
| 101 | c214_ryspath_iq4kss |
0.8575 | 0.8786 | 0.8300 | 0.8512 | 16.48 GB | 0 | rys_path=iq4_kss | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at iq4_kss.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_kss |
| 102 | c190_faqkv_iq4kt |
0.8559 | 0.8786 | 0.8263 | 0.8492 | 16.46 GB | 0 | fa_qkv=iq4_kt | Mixed manual probe: full-attention Q/K/V tensors at iq4_kt.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=iq4_kt |
| 103 | c242_bndall_mxfp4 |
0.8559 | 0.8786 | 0.8263 | 0.8492 | 16.51 GB | 0 | boundary_all=mxfp4 | Mixed manual probe: RYS entry, exit, and duplicate-boundary attention/SSM/FFN tensors at mxfp4.rules^blk\.(15|19|20|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=mxfp4 |
| 104 | c348_srcsig_mxfp4 |
0.8559 | 0.8786 | 0.8263 | 0.8492 | 16.54 GB | 0 | src_signal=mxfp4 | Mixed manual probe: source-side RYS 15..19 linear-attention and SSM signal path at mxfp4.rules^blk\.(15|16|17|18|19)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=mxfp4 |
| 105 | c450_ryssig_q5kr4 |
0.8559 | 0.8786 | 0.8263 | 0.8492 | 16.64 GB | 0 | rys_signal=q5_k_r4 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at q5_k_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q5_k_r4 |
| 106 | c277_lateffn_iq5k |
0.8559 | 0.8786 | 0.8263 | 0.8492 | 16.64 GB | 0 | late_ffn=iq5_k | Mixed manual probe: late FFN gate/down control tensors at iq5_k.rules^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq5_k |
| 107 | c590_fa_iq5ks_lateffn_iq4kss |
0.8559 | 0.8786 | 0.8263 | 0.8492 | 16.68 GB | 0 | fa=iq5_ks, late_ffn=iq4_kss | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_ks; late FFN gate/down control tensors at iq4_kss.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq4_kss |
| 108 | c492_lateattn_q8kv |
0.8559 | 0.8786 | 0.8263 | 0.8492 | 16.78 GB | 0 | late_attn=q8_KV | Mixed manual probe: late full-attention control tensors at q8_KV.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_KV |
| 109 | c429_ryspath_q60r4 |
0.8559 | 0.8786 | 0.8263 | 0.8492 | 16.83 GB | 0 | rys_path=q6_0_r4 | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at q6_0_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q6_0_r4 |
| 110 | c494_fa_iq4ks_outtok_iq4kss |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.00 GB | 0 | fa=iq4_ks, output_token=iq4_kss | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_ks; token embedding and output tensors at iq4_kss.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_ks ^output\.weight=iq4_kss ^token_embd\.weight=iq4_kss |
| 111 | c734_fa_iq3kt_lateffn_iq4kss |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.18 GB | 0 | fa=iq3_kt, late_ffn=iq4_kss | Mixed manual probe: all full-attention Q/K/V/O tensors at iq3_kt; late FFN gate/down control tensors at iq4_kss.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq3_kt ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq4_kss |
| 112 | c171_fa_iq3kt |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.22 GB | 0 | fa=iq3_kt | Mixed manual probe: all full-attention Q/K/V/O tensors at iq3_kt.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq3_kt |
| 113 | c306_outtok_iq4k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.23 GB | 0 | output_token=iq4_k | Mixed manual probe: token embedding and output tensors at iq4_k.rules^output\.weight=iq4_k ^token_embd\.weight=iq4_k |
| 114 | c471_outtok_iq4kr4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.23 GB | 0 | output_token=iq4_k_r4 | Mixed manual probe: token embedding and output tensors at iq4_k_r4.rules^output\.weight=iq4_k_r4 ^token_embd\.weight=iq4_k_r4 |
| 115 | c241_bndall_iq3kt |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.30 GB | 0 | boundary_all=iq3_kt | Mixed manual probe: RYS entry, exit, and duplicate-boundary attention/SSM/FFN tensors at iq3_kt.rules^blk\.(15|19|20|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=iq3_kt |
| 116 | c191_faqkv_iq3kt |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.32 GB | 0 | fa_qkv=iq3_kt | Mixed manual probe: full-attention Q/K/V tensors at iq3_kt.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=iq3_kt |
| 117 | c315_out_iq5ks |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.35 GB | 0 | output_only=iq5_ks | Mixed manual probe: output tensor at iq5_ks.rules^output\.weight=iq5_ks |
| 118 | c692_fa_iq4kt_bndffn_iq4kss |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.36 GB | 0 | fa=iq4_kt, boundary_ffn=iq4_kss | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kt; RYS entry, exit, and duplicate-boundary FFN tensors at iq4_kss.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kt ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=iq4_kss |
| 119 | c221_ryspath_iq3kt |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.36 GB | 0 | rys_path=iq3_kt | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at iq3_kt.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq3_kt |
| 120 | c251_bndffn_iq3kt |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.37 GB | 0 | boundary_ffn=iq3_kt | Mixed manual probe: RYS entry, exit, and duplicate-boundary FFN tensors at iq3_kt.rules^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=iq3_kt |
| 121 | c505_fa_iq4ks_rysffn_iq4ks |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.40 GB | 0 | fa=iq4_ks, rys_ffn=iq4_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_ks; RYS 15..24 FFN tensors at iq4_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_ks ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=iq4_ks |
| 122 | c164_fa_iq4kss |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.43 GB | 0 | fa=iq4_kss | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kss.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kss |
| 123 | c170_fa_iq4kt |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.43 GB | 0 | fa=iq4_kt | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kt.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kt |
| 124 | c524_fa_iq4ks_finalffn_iq4kss |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.44 GB | 0 | fa=iq4_ks, final_ffn=iq4_kss | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_ks; final FFN tensors at iq4_kss.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_ks ^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=iq4_kss |
| 125 | c231_ryssig_iq3kt |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.44 GB | 0 | rys_signal=iq3_kt | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at iq3_kt.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq3_kt |
| 126 | c271_lateattn_iq3kt |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.46 GB | 0 | late_attn=iq3_kt | Mixed manual probe: late full-attention control tensors at iq3_kt.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq3_kt |
| 127 | c470_outtok_iq5ksr4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.47 GB | 0 | output_token=iq5_ks_r4 | Mixed manual probe: token embedding and output tensors at iq5_ks_r4.rules^output\.weight=iq5_ks_r4 ^token_embd\.weight=iq5_ks_r4 |
| 128 | c259_rysffn_iq4xs |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.47 GB | 0 | rys_ffn=iq4_xs | Mixed manual probe: RYS 15..24 FFN tensors at iq4_xs.rules^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=iq4_xs |
| 129 | c290_latectl_iq4kt |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.47 GB | 0 | late_control=iq4_kt | Mixed manual probe: late attention plus FFN gate/down control tensors at iq4_kt.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kt ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq4_kt |
| 130 | c180_faqk_iq4kt |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.48 GB | 0 | fa_qk=iq4_kt | Mixed manual probe: full-attention Q/K tensors at iq4_kt.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=iq4_kt |
| 131 | c220_ryspath_iq4kt |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.48 GB | 0 | rys_path=iq4_kt | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at iq4_kt.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_kt |
| 132 | c250_bndffn_iq4kt |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.49 GB | 0 | boundary_ffn=iq4_kt | Mixed manual probe: RYS entry, exit, and duplicate-boundary FFN tensors at iq4_kt.rules^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=iq4_kt |
| 133 | c353_dupsig_iq3kt |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.50 GB | 0 | dup_signal=iq3_kt | Mixed manual probe: duplicated-side RYS 20..24 linear-attention and SSM signal path at iq3_kt.rules^blk\.(20|21|22|23|24)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq3_kt |
| 134 | c189_faqkv_iq4xs |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.50 GB | 0 | fa_qkv=iq4_xs | Mixed manual probe: full-attention Q/K/V tensors at iq4_xs.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=iq4_xs |
| 135 | c183_faqkv_iq4ks |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.50 GB | 0 | fa_qkv=iq4_ks | Mixed manual probe: full-attention Q/K/V tensors at iq4_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=iq4_ks |
| 136 | c233_bndall_iq4ks |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.51 GB | 0 | boundary_all=iq4_ks | Mixed manual probe: RYS entry, exit, and duplicate-boundary attention/SSM/FFN tensors at iq4_ks.rules^blk\.(15|19|20|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=iq4_ks |
| 137 | c280_lateffn_iq4kt |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.51 GB | 0 | late_ffn=iq4_kt | Mixed manual probe: late FFN gate/down control tensors at iq4_kt.rules^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq4_kt |
| 138 | c224_ryssig_iq4kss |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.51 GB | 0 | rys_signal=iq4_kss | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at iq4_kss.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_kss |
| 139 | c230_ryssig_iq4kt |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.51 GB | 0 | rys_signal=iq4_kt | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at iq4_kt.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_kt |
| 140 | c289_latectl_iq4xs |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.51 GB | 0 | late_control=iq4_xs | Mixed manual probe: late attention plus FFN gate/down control tensors at iq4_xs.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_xs ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq4_xs |
| 141 | c332_tok_mxfp4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.52 GB | 0 | token_only=mxfp4 | Mixed manual probe: token embedding tensor at mxfp4.rules^token_embd\.weight=mxfp4 |
| 142 | c323_tok_iq4ks |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.52 GB | 0 | token_only=iq4_ks | Mixed manual probe: token embedding tensor at iq4_ks.rules^token_embd\.weight=iq4_ks |
| 143 | c182_faqk_mxfp4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.52 GB | 0 | fa_qk=mxfp4 | Mixed manual probe: full-attention Q/K tensors at mxfp4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=mxfp4 |
| 144 | c422_ryspath_iq4xsr8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.52 GB | 0 | rys_path=iq4_xs_r8 | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at iq4_xs_r8.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_xs_r8 |
| 145 | c173_faqk_iq4ks |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.52 GB | 0 | fa_qk=iq4_ks | Mixed manual probe: full-attention Q/K tensors at iq4_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=iq4_ks |
| 146 | c384_faqk_iq4ksr4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.52 GB | 0 | fa_qk=iq4_ks_r4 | Mixed manual probe: full-attention Q/K tensors at iq4_ks_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=iq4_ks_r4 |
| 147 | c418_ryspath_iq4ksr4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.52 GB | 0 | rys_path=iq4_ks_r4 | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at iq4_ks_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_ks_r4 |
| 148 | c358_srcpath_iq4kt |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.52 GB | 0 | src_path=iq4_kt | Mixed manual probe: source-side RYS 15..19 attention, linear-attention, and SSM path at iq4_kt.rules^blk\.(15|16|17|18|19)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_kt |
| 149 | c270_lateattn_iq4kt |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.52 GB | 0 | late_attn=iq4_kt | Mixed manual probe: late full-attention control tensors at iq4_kt.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kt |
| 150 | c243_bndffn_iq4ks |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.52 GB | 0 | boundary_ffn=iq4_ks | Mixed manual probe: RYS entry, exit, and duplicate-boundary FFN tensors at iq4_ks.rules^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=iq4_ks |
| 151 | c334_earlyattn_iq4kss |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.53 GB | 0 | early_attn=iq4_kss | Mixed manual probe: early full-attention routing tensors at iq4_kss.rules^blk\.(3|7|11|15)\.attn_(q|k|v|output)\.weight=iq4_kss |
| 152 | c340_earlyattn_iq4kt |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.53 GB | 0 | early_attn=iq4_kt | Mixed manual probe: early full-attention routing tensors at iq4_kt.rules^blk\.(3|7|11|15)\.attn_(q|k|v|output)\.weight=iq4_kt |
| 153 | c299_finalffn_iq4xs |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.53 GB | 0 | final_ffn=iq4_xs | Mixed manual probe: final FFN tensors at iq4_xs.rules^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=iq4_xs |
| 154 | c229_ryssig_iq4xs |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.53 GB | 0 | rys_signal=iq4_xs | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at iq4_xs.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_xs |
| 155 | c350_dupsig_iq4kss |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.53 GB | 0 | dup_signal=iq4_kss | Mixed manual probe: duplicated-side RYS 20..24 linear-attention and SSM signal path at iq4_kss.rules^blk\.(20|21|22|23|24)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_kss |
| 156 | c435_ryssig_iq4ksr4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.53 GB | 0 | rys_signal=iq4_ks_r4 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at iq4_ks_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_ks_r4 |
| 157 | c269_lateattn_iq4xs |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.54 GB | 0 | late_attn=iq4_xs | Mixed manual probe: late full-attention control tensors at iq4_xs.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_xs |
| 158 | c366_duppath_mxfp4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.54 GB | 0 | dup_path=mxfp4 | Mixed manual probe: duplicated-side RYS 20..24 attention, linear-attention, and SSM path at mxfp4.rules^blk\.(20|21|22|23|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=mxfp4 |
| 159 | c333_earlyattn_iq4ks |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.54 GB | 0 | early_attn=iq4_ks | Mixed manual probe: early full-attention routing tensors at iq4_ks.rules^blk\.(3|7|11|15)\.attn_(q|k|v|output)\.weight=iq4_ks |
| 160 | c196_fav_iq4k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.54 GB | 0 | fa_v=iq4_k | Mixed manual probe: full-attention V tensors at iq4_k.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=iq4_k |
| 161 | c354_dupsig_mxfp4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.54 GB | 0 | dup_signal=mxfp4 | Mixed manual probe: duplicated-side RYS 20..24 linear-attention and SSM signal path at mxfp4.rules^blk\.(20|21|22|23|24)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=mxfp4 |
| 162 | c349_dupsig_iq4ks |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.54 GB | 0 | dup_signal=iq4_ks | Mixed manual probe: duplicated-side RYS 20..24 linear-attention and SSM signal path at iq4_ks.rules^blk\.(20|21|22|23|24)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_ks |
| 163 | c307_outtok_iq5k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.54 GB | 0 | output_token=iq5_k | Mixed manual probe: token embedding and output tensors at iq5_k.rules^output\.weight=iq5_k ^token_embd\.weight=iq5_k |
| 164 | c481_outtok_q50r4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.54 GB | 0 | output_token=q5_0_r4 | Mixed manual probe: token embedding and output tensors at q5_0_r4.rules^output\.weight=q5_0_r4 ^token_embd\.weight=q5_0_r4 |
| 165 | c484_outtok_q5kr4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.54 GB | 0 | output_token=q5_k_r4 | Mixed manual probe: token embedding and output tensors at q5_k_r4.rules^output\.weight=q5_k_r4 ^token_embd\.weight=q5_k_r4 |
| 166 | c465_latectl_q40r8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.55 GB | 0 | late_control=q4_0_r8 | Mixed manual probe: late attention plus FFN gate/down control tensors at q4_0_r8.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q4_0_r8 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q4_0_r8 |
| 167 | c216_ryspath_iq4k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.55 GB | 0 | rys_path=iq4_k | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at iq4_k.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_k |
| 168 | c236_bndall_iq4k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.55 GB | 0 | boundary_all=iq4_k | Mixed manual probe: RYS entry, exit, and duplicate-boundary attention/SSM/FFN tensors at iq4_k.rules^blk\.(15|19|20|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=iq4_k |
| 169 | c176_faqk_iq4k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.55 GB | 0 | fa_qk=iq4_k | Mixed manual probe: full-attention Q/K tensors at iq4_k.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=iq4_k |
| 170 | c197_fav_iq5k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.55 GB | 0 | fa_v=iq5_k | Mixed manual probe: full-attention V tensors at iq5_k.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=iq5_k |
| 171 | c386_faqk_iq4kr4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.55 GB | 0 | fa_qk=iq4_k_r4 | Mixed manual probe: full-attention Q/K tensors at iq4_k_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=iq4_k_r4 |
| 172 | c398_faqk_q4kr4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.55 GB | 0 | fa_qk=q4_k_r4 | Mixed manual probe: full-attention Q/K tensors at q4_k_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=q4_k_r4 |
| 173 | c440_ryssig_iq4nlr4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.55 GB | 0 | rys_signal=iq4_nl_r4 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at iq4_nl_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_nl_r4 |
| 174 | c449_ryssig_q4kr4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.55 GB | 0 | rys_signal=q4_k_r4 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at q4_k_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q4_k_r4 |
| 175 | c318_out_iq6k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.56 GB | 0 | output_only=iq6_k | Mixed manual probe: output tensor at iq6_k.rules^output\.weight=iq6_k |
| 176 | c198_fav_iq6k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.57 GB | 0 | fa_v=iq6_k | Mixed manual probe: full-attention V tensors at iq6_k.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=iq6_k |
| 177 | c345_srcsig_iq5ks |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.59 GB | 0 | src_signal=iq5_ks | Mixed manual probe: source-side RYS 15..19 linear-attention and SSM signal path at iq5_ks.rules^blk\.(15|16|17|18|19)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq5_ks |
| 178 | c335_earlyattn_iq5ks |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.59 GB | 0 | early_attn=iq5_ks | Mixed manual probe: early full-attention routing tensors at iq5_ks.rules^blk\.(3|7|11|15)\.attn_(q|k|v|output)\.weight=iq5_ks |
| 179 | c627_fa_iq4k_lateffn_iq5ks |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.61 GB | 0 | fa=iq4_k, late_ffn=iq5_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_k; late FFN gate/down control tensors at iq5_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_k ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq5_ks |
| 180 | c633_fa_iq4k_finalffn_iq5ks |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.62 GB | 0 | fa=iq4_k, final_ffn=iq5_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_k; final FFN tensors at iq5_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_k ^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=iq5_ks |
| 181 | c225_ryssig_iq5ks |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.62 GB | 0 | rys_signal=iq5_ks | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at iq5_ks.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq5_ks |
| 182 | c436_ryssig_iq5ksr4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.62 GB | 0 | rys_signal=iq5_ks_r4 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at iq5_ks_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq5_ks_r4 |
| 183 | c621_fa_iq4k_bndffn_iq5ks |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.64 GB | 0 | fa=iq4_k, boundary_ffn=iq5_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_k; RYS entry, exit, and duplicate-boundary FFN tensors at iq5_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_k ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=iq5_ks |
| 184 | c338_earlyattn_iq6k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.66 GB | 0 | early_attn=iq6_k | Mixed manual probe: early full-attention routing tensors at iq6_k.rules^blk\.(3|7|11|15)\.attn_(q|k|v|output)\.weight=iq6_k |
| 185 | c453_latectl_iq5ksr4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.67 GB | 0 | late_control=iq5_ks_r4 | Mixed manual probe: late attention plus FFN gate/down control tensors at iq5_ks_r4.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks_r4 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq5_ks_r4 |
| 186 | c385_faqk_iq5ksr4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.67 GB | 0 | fa_qk=iq5_ks_r4 | Mixed manual probe: full-attention Q/K tensors at iq5_ks_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=iq5_ks_r4 |
| 187 | c421_ryspath_iq5kr4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.69 GB | 0 | rys_path=iq5_k_r4 | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at iq5_k_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq5_k_r4 |
| 188 | c187_faqkv_iq5k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.71 GB | 0 | fa_qkv=iq5_k | Mixed manual probe: full-attention Q/K/V tensors at iq5_k.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=iq5_k |
| 189 | c387_faqk_iq5kr4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.71 GB | 0 | fa_qk=iq5_k_r4 | Mixed manual probe: full-attention Q/K tensors at iq5_k_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=iq5_k_r4 |
| 190 | c327_tok_iq5k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.71 GB | 0 | token_only=iq5_k | Mixed manual probe: token embedding tensor at iq5_k.rules^token_embd\.weight=iq5_k |
| 191 | c368_fa_iq5ksr4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.72 GB | 0 | fa=iq5_ks_r4 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_ks_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks_r4 |
| 192 | c237_bndall_iq5k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.74 GB | 0 | boundary_all=iq5_k | Mixed manual probe: RYS entry, exit, and duplicate-boundary attention/SSM/FFN tensors at iq5_k.rules^blk\.(15|19|20|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=iq5_k |
| 193 | c228_ryssig_iq6k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.74 GB | 0 | rys_signal=iq6_k | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at iq6_k.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq6_k |
| 194 | c278_lateffn_iq6k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.74 GB | 0 | late_ffn=iq6_k | Mixed manual probe: late FFN gate/down control tensors at iq6_k.rules^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq6_k |
| 195 | c676_fa_iq4kt_outtok_q6k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.75 GB | 0 | fa=iq4_kt, output_token=q6_K | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kt; token embedding and output tensors at q6_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kt ^output\.weight=q6_K ^token_embd\.weight=q6_K |
| 196 | c539_fa_iq4kss_ryssig_q8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.77 GB | 0 | fa=iq4_kss, rys_signal=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kss; RYS 15..24 linear-attention and SSM signal path at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kss ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 197 | c428_ryspath_q60 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.83 GB | 0 | rys_path=q6_0 | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at q6_0.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q6_0 |
| 198 | c248_bndffn_iq6k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.84 GB | 0 | boundary_ffn=iq6_k | Mixed manual probe: RYS entry, exit, and duplicate-boundary FFN tensors at iq6_k.rules^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=iq6_k |
| 199 | c434_ryspath_q6kr4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.84 GB | 0 | rys_path=q6_k_r4 | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at q6_k_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q6_k_r4 |
| 200 | c218_ryspath_iq6k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.85 GB | 0 | rys_path=iq6_k | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at iq6_k.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq6_k |
| 201 | c444_ryssig_q8kvr8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.86 GB | 0 | rys_signal=q8_KV_r8 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at q8_KV_r8.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_KV_r8 |
| 202 | c463_latectl_q60r4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.86 GB | 0 | late_control=q6_0_r4 | Mixed manual probe: late attention plus FFN gate/down control tensors at q6_0_r4.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_0_r4 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q6_0_r4 |
| 203 | c395_faqk_q60r4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.86 GB | 0 | fa_qk=q6_0_r4 | Mixed manual probe: full-attention Q/K tensors at q6_0_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=q6_0_r4 |
| 204 | c480_outtok_q60r4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.86 GB | 0 | output_token=q6_0_r4 | Mixed manual probe: token embedding and output tensors at q6_0_r4.rules^output\.weight=q6_0_r4 ^token_embd\.weight=q6_0_r4 |
| 205 | c443_ryssig_q8kr8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.86 GB | 0 | rys_signal=q8_k_r8 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at q8_k_r8.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_k_r8 |
| 206 | c411_faqkv_q60 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.87 GB | 0 | fa_qkv=q6_0 | Mixed manual probe: full-attention Q/K/V tensors at q6_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=q6_0 |
| 207 | c412_faqkv_q60r4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.87 GB | 0 | fa_qkv=q6_0_r4 | Mixed manual probe: full-attention Q/K/V tensors at q6_0_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=q6_0_r4 |
| 208 | c288_latectl_iq6k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.88 GB | 0 | late_control=iq6_k | Mixed manual probe: late attention plus FFN gate/down control tensors at iq6_k.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq6_k ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq6_k |
| 209 | c417_faqkv_q6kr4 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.88 GB | 0 | fa_qkv=q6_k_r4 | Mixed manual probe: full-attention Q/K/V tensors at q6_k_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=q6_k_r4 |
| 210 | c442_ryssig_q8r8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.90 GB | 0 | rys_signal=q8_0_r8 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at q8_0_r8.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0_r8 |
| 211 | c308_outtok_iq6k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.90 GB | 0 | output_token=iq6_k | Mixed manual probe: token embedding and output tensors at iq6_k.rules^output\.weight=iq6_k ^token_embd\.weight=iq6_k |
| 212 | c35_bndall_q6k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.94 GB | 0 | boundary_all=q6_K | Mixed manual probe: RYS entry, exit, and duplicate-boundary attention/SSM/FFN tensors at q6_K.rules^blk\.(15|19|20|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=q6_K |
| 213 | c238_bndall_iq6k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 16.95 GB | 0 | boundary_all=iq6_k | Mixed manual probe: RYS entry, exit, and duplicate-boundary attention/SSM/FFN tensors at iq6_k.rules^blk\.(15|19|20|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=iq6_k |
| 214 | c427_ryspath_q8kvr8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 17.04 GB | 0 | rys_path=q8_KV_r8 | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at q8_KV_r8.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_KV_r8 |
| 215 | c461_latectl_q8kvr8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 17.09 GB | 0 | late_control=q8_KV_r8 | Mixed manual probe: late attention plus FFN gate/down control tensors at q8_KV_r8.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_KV_r8 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_KV_r8 |
| 216 | c460_latectl_q8kr8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 17.10 GB | 0 | late_control=q8_k_r8 | Mixed manual probe: late attention plus FFN gate/down control tensors at q8_k_r8.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_k_r8 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_k_r8 |
| 217 | c688_fa_iq4kt_rysffn_q6k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 17.12 GB | 0 | fa=iq4_kt, rys_ffn=q6_K | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kt; RYS 15..24 FFN tensors at q6_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kt ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q6_K |
| 218 | c599_fa_iq5ks_finalffn_q8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 17.12 GB | 0 | fa=iq5_ks, final_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_ks; final FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks ^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=q8_0 |
| 219 | c409_faqkv_q8kr8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 17.13 GB | 0 | fa_qkv=q8_k_r8 | Mixed manual probe: full-attention Q/K/V tensors at q8_k_r8.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=q8_k_r8 |
| 220 | c408_faqkv_q8r8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 17.20 GB | 0 | fa_qkv=q8_0_r8 | Mixed manual probe: full-attention Q/K/V tensors at q8_0_r8.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=q8_0_r8 |
| 221 | c03_habit_boundary_q8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 17.30 GB | 0 | ^blk\.(15|19|20|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=q8_0 | Protect the RYS entry/exit boundaries and duplicated boundary layers with Q8_0 across attention, SSM, and FFN tensors.rules^blk\.(15|19|20|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=q8_0 |
| 222 | c45_lateattn_bf16 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 17.31 GB | 0 | late_attn=bf16 | Mixed manual probe: late full-attention control tensors at bf16.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 |
| 223 | c478_outtok_q8kvr8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 17.34 GB | 0 | output_token=q8_KV_r8 | Mixed manual probe: token embedding and output tensors at q8_KV_r8.rules^output\.weight=q8_KV_r8 ^token_embd\.weight=q8_KV_r8 |
| 224 | c580_fa_iq5ks_rysffn_q6k |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 17.41 GB | 0 | fa=iq5_ks, rys_ffn=q6_K | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_ks; RYS 15..24 FFN tensors at q6_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q6_K |
| 225 | c477_outtok_q8kr8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 17.43 GB | 0 | output_token=q8_k_r8 | Mixed manual probe: token embedding and output tensors at q8_k_r8.rules^output\.weight=q8_k_r8 ^token_embd\.weight=q8_k_r8 |
| 226 | c702_fa_iq4kt_lateffn_bf16 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 17.45 GB | 0 | fa=iq4_kt, late_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kt; late FFN gate/down control tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kt ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=bf16 |
| 227 | c476_outtok_q8r8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 17.50 GB | 0 | output_token=q8_0_r8 | Mixed manual probe: token embedding and output tensors at q8_0_r8.rules^output\.weight=q8_0_r8 ^token_embd\.weight=q8_0_r8 |
| 228 | c689_fa_iq4kt_rysffn_q8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 17.76 GB | 0 | fa=iq4_kt, rys_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kt; RYS 15..24 FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kt ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q8_0 |
| 229 | c89_fa_q8_ryssig_q8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 17.83 GB | 0 | fa=q8_0, rys_signal=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; RYS 15..24 linear-attention and SSM signal path at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 230 | c672_fa_iq5k_finalffn_bf16 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 17.93 GB | 0 | fa=iq5_k, final_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_k; final FFN tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_k ^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=bf16 |
| 231 | c114_fao_bf16_faqk_q8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 17.98 GB | 0 | fa_o=bf16, fa_qk=q8_0 | Mixed manual probe: full-attention output tensors at bf16; full-attention Q/K tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=bf16 ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=q8_0 |
| 232 | c15_faqk_bf16 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 18.32 GB | 0 | fa_qk=bf16 | Mixed manual probe: full-attention Q/K tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=bf16 |
| 233 | c87_fa_q8_outtok_q8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 18.43 GB | 0 | fa=q8_0, output_token=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; token embedding and output tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 |
| 234 | c101_fa_bf16_ryssig_q8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 19.60 GB | 0 | fa=bf16, rys_signal=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; RYS 15..24 linear-attention and SSM signal path at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 235 | c534_fa_iq4kss_outtok_bf16 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 19.75 GB | 0 | fa=iq4_kss, output_token=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kss; token embedding and output tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kss ^output\.weight=bf16 ^token_embd\.weight=bf16 |
| 236 | c105_fa_bf16_bndffn_q8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 19.79 GB | 0 | fa=bf16, boundary_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; RYS entry, exit, and duplicate-boundary FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 |
| 237 | c99_fa_bf16_outtok_q8 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 20.20 GB | 0 | fa=bf16, output_token=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; token embedding and output tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 |
| 238 | c102_fa_bf16_ryssig_bf16 |
0.8539 | 0.8786 | 0.8219 | 0.8467 | 20.26 GB | 0 | fa=bf16, rys_signal=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; RYS 15..24 linear-attention and SSM signal path at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 |
| 239 | c319_out_iq4xs |
0.8539 | 0.9143 | 0.7752 | 0.8360 | 16.19 GB | 0 | output_only=iq4_xs | Mixed manual probe: output tensor at iq4_xs.rules^output\.weight=iq4_xs |
| 240 | c644_fa_iq5k_ryssig_iq4kss |
0.8539 | 0.9143 | 0.7752 | 0.8360 | 16.74 GB | 0 | fa=iq5_k, rys_signal=iq4_kss | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_k; RYS 15..24 linear-attention and SSM signal path at iq4_kss.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_k ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_kss |
| 241 | c142_fa_q8_outtok_q8_lateffn_q8 |
0.8539 | 0.9143 | 0.7752 | 0.8360 | 18.79 GB | 0 | fa=q8_0, output_token=q8_0, late_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; token embedding and output tensors at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 242 | c148_fa_bf16_outtok_q8_lateffn_q8 |
0.8539 | 0.9143 | 0.7752 | 0.8360 | 20.56 GB | 0 | fa=bf16, output_token=q8_0, late_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; token embedding and output tensors at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 243 | c211_fao_iq3kt |
0.8524 | 0.8929 | 0.7996 | 0.8404 | 16.46 GB | 0 | fa_o=iq3_kt | Mixed manual probe: full-attention output tensors at iq3_kt.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=iq3_kt |
| 244 | c466_latectl_q4kr4 |
0.8520 | 0.8786 | 0.8174 | 0.8442 | 16.55 GB | 0 | late_control=q4_k_r4 | Mixed manual probe: late attention plus FFN gate/down control tensors at q4_k_r4.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q4_k_r4 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q4_k_r4 |
| 245 | c298_finalffn_iq6k |
0.8520 | 0.8786 | 0.8174 | 0.8442 | 16.77 GB | 0 | final_ffn=iq6_k | Mixed manual probe: final FFN tensors at iq6_k.rules^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=iq6_k |
| 246 | c722_fa_iq3kt_rysffn_iq4kss |
0.8497 | 0.8929 | 0.7935 | 0.8370 | 16.05 GB | 0 | fa=iq3_kt, rys_ffn=iq4_kss | Mixed manual probe: all full-attention Q/K/V/O tensors at iq3_kt; RYS 15..24 FFN tensors at iq4_kss.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq3_kt ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=iq4_kss |
| 247 | c262_rysffn_mxfp4 |
0.8471 | 0.8786 | 0.8061 | 0.8378 | 16.47 GB | 0 | rys_ffn=mxfp4 | Mixed manual probe: RYS 15..24 FFN tensors at mxfp4.rules^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=mxfp4 |
| 248 | c309_outtok_iq4xs |
0.8462 | 0.8786 | 0.8041 | 0.8367 | 16.15 GB | 0 | output_token=iq4_xs | Mixed manual probe: token embedding and output tensors at iq4_xs.rules^output\.weight=iq4_xs ^token_embd\.weight=iq4_xs |
| 249 | c482_outtok_q40r8 |
0.8462 | 0.8786 | 0.8041 | 0.8367 | 16.23 GB | 0 | output_token=q4_0_r8 | Mixed manual probe: token embedding and output tensors at q4_0_r8.rules^output\.weight=q4_0_r8 ^token_embd\.weight=q4_0_r8 |
| 250 | c424_ryspath_iq3kr4 |
0.8462 | 0.8786 | 0.8041 | 0.8367 | 16.40 GB | 0 | rys_path=iq3_k_r4 | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at iq3_k_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq3_k_r4 |
| 251 | c302_finalffn_mxfp4 |
0.8462 | 0.8786 | 0.8041 | 0.8367 | 16.53 GB | 0 | final_ffn=mxfp4 | Mixed manual probe: final FFN tensors at mxfp4.rules^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=mxfp4 |
| 252 | c282_lateffn_mxfp4 |
0.8462 | 0.8786 | 0.8041 | 0.8367 | 16.53 GB | 0 | late_ffn=mxfp4 | Mixed manual probe: late FFN gate/down control tensors at mxfp4.rules^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=mxfp4 |
| 253 | c232_ryssig_mxfp4 |
0.8462 | 0.8786 | 0.8041 | 0.8367 | 16.53 GB | 0 | rys_signal=mxfp4 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at mxfp4.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=mxfp4 |
| 254 | c381_fa_q4kr4 |
0.8462 | 0.8786 | 0.8041 | 0.8367 | 16.54 GB | 0 | fa=q4_k_r4 | Mixed manual probe: all full-attention Q/K/V/O tensors at q4_k_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q4_k_r4 |
| 255 | c195_fav_iq5ks |
0.8462 | 0.8786 | 0.8041 | 0.8367 | 16.55 GB | 0 | fa_v=iq5_ks | Mixed manual probe: full-attention V tensors at iq5_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=iq5_ks |
| 256 | c419_ryspath_iq5ksr4 |
0.8462 | 0.8786 | 0.8041 | 0.8367 | 16.66 GB | 0 | rys_path=iq5_ks_r4 | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at iq5_ks_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq5_ks_r4 |
| 257 | c321_out_iq3kt |
0.8460 | 0.8786 | 0.8035 | 0.8364 | 16.01 GB | 0 | output_only=iq3_kt | Mixed manual probe: output tensor at iq3_kt.rules^output\.weight=iq3_kt |
| 258 | c346_srcsig_iq4kt |
0.8460 | 0.8786 | 0.8035 | 0.8364 | 16.53 GB | 0 | src_signal=iq4_kt | Mixed manual probe: source-side RYS 15..19 linear-attention and SSM signal path at iq4_kt.rules^blk\.(15|16|17|18|19)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_kt |
| 259 | c295_finalffn_iq5ks |
0.8460 | 0.8786 | 0.8035 | 0.8364 | 16.63 GB | 0 | final_ffn=iq5_ks | Mixed manual probe: final FFN tensors at iq5_ks.rules^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=iq5_ks |
| 260 | c287_latectl_iq5k |
0.8452 | 0.8786 | 0.8019 | 0.8354 | 16.71 GB | 0 | late_control=iq5_k | Mixed manual probe: late attention plus FFN gate/down control tensors at iq5_k.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_k ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq5_k |
| 261 | c591_fa_iq5ks_lateffn_iq5ks |
0.8452 | 0.8786 | 0.8019 | 0.8354 | 16.79 GB | 0 | fa=iq5_ks, late_ffn=iq5_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_ks; late FFN gate/down control tensors at iq5_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq5_ks |
| 262 | c393_faqk_q8kvr8 |
0.8452 | 0.8786 | 0.8019 | 0.8354 | 17.09 GB | 0 | fa_qk=q8_KV_r8 | Mixed manual probe: full-attention Q/K tensors at q8_KV_r8.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=q8_KV_r8 |
| 263 | c475_outtok_iq3kr4 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 15.89 GB | 0 | output_token=iq3_k_r4 | Mixed manual probe: token embedding and output tensors at iq3_k_r4.rules^output\.weight=iq3_k_r4 ^token_embd\.weight=iq3_k_r4 |
| 264 | c529_fa_iq4kss_outtok_iq4ks |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.02 GB | 0 | fa=iq4_kss, output_token=iq4_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kss; token embedding and output tensors at iq4_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kss ^output\.weight=iq4_ks ^token_embd\.weight=iq4_ks |
| 265 | c473_outtok_iq4xsr8 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.15 GB | 0 | output_token=iq4_xs_r8 | Mixed manual probe: token embedding and output tensors at iq4_xs_r8.rules^output\.weight=iq4_xs_r8 ^token_embd\.weight=iq4_xs_r8 |
| 266 | c303_outtok_iq4ks |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.15 GB | 0 | output_token=iq4_ks | Mixed manual probe: token embedding and output tensors at iq4_ks.rules^output\.weight=iq4_ks ^token_embd\.weight=iq4_ks |
| 267 | c313_out_iq4ks |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.19 GB | 0 | output_only=iq4_ks | Mixed manual probe: output tensor at iq4_ks.rules^output\.weight=iq4_ks |
| 268 | c316_out_iq4k |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.23 GB | 0 | output_only=iq4_k | Mixed manual probe: output tensor at iq4_k.rules^output\.weight=iq4_k |
| 269 | c483_outtok_q4kr4 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.23 GB | 0 | output_token=q4_k_r4 | Mixed manual probe: token embedding and output tensors at q4_k_r4.rules^output\.weight=q4_k_r4 ^token_embd\.weight=q4_k_r4 |
| 270 | c741_fa_iq3kt_finalffn_iq5ks |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.30 GB | 0 | fa=iq3_kt, final_ffn=iq5_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq3_kt; final FFN tensors at iq5_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq3_kt ^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=iq5_ks |
| 271 | c331_tok_iq3kt |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.34 GB | 0 | token_only=iq3_kt | Mixed manual probe: token embedding tensor at iq3_kt.rules^token_embd\.weight=iq3_kt |
| 272 | c291_latectl_iq3kt |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.34 GB | 0 | late_control=iq3_kt | Mixed manual probe: late attention plus FFN gate/down control tensors at iq3_kt.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq3_kt ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq3_kt |
| 273 | c181_faqk_iq3kt |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.34 GB | 0 | fa_qk=iq3_kt | Mixed manual probe: full-attention Q/K tensors at iq3_kt.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=iq3_kt |
| 274 | c554_fa_iq4kss_lateffn_iq4kss |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.38 GB | 0 | fa=iq4_kss, late_ffn=iq4_kss | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kss; late FFN gate/down control tensors at iq4_kss.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kss ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq4_kss |
| 275 | c698_fa_iq4kt_lateffn_iq4kss |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.38 GB | 0 | fa=iq4_kt, late_ffn=iq4_kss | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kt; late FFN gate/down control tensors at iq4_kss.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kt ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq4_kss |
| 276 | c680_fa_iq4kt_ryssig_iq4kss |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.38 GB | 0 | fa=iq4_kt, rys_signal=iq4_kss | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kt; RYS 15..24 linear-attention and SSM signal path at iq4_kss.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kt ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_kss |
| 277 | c260_rysffn_iq4kt |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.39 GB | 0 | rys_ffn=iq4_kt | Mixed manual probe: RYS 15..24 FFN tensors at iq4_kt.rules^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=iq4_kt |
| 278 | c559_fa_iq4kss_finalffn_iq4ks |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.40 GB | 0 | fa=iq4_kss, final_ffn=iq4_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kss; final FFN tensors at iq4_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kss ^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=iq4_ks |
| 279 | c301_finalffn_iq3kt |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.42 GB | 0 | final_ffn=iq3_kt | Mixed manual probe: final FFN tensors at iq3_kt.rules^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=iq3_kt |
| 280 | c511_fa_iq4ks_bndffn_iq4ks |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.45 GB | 0 | fa=iq4_ks, boundary_ffn=iq4_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_ks; RYS entry, exit, and duplicate-boundary FFN tensors at iq4_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_ks ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=iq4_ks |
| 281 | c365_duppath_iq3kt |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.46 GB | 0 | dup_path=iq3_kt | Mixed manual probe: duplicated-side RYS 20..24 attention, linear-attention, and SSM path at iq3_kt.rules^blk\.(20|21|22|23|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq3_kt |
| 282 | c305_outtok_iq5ks |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.47 GB | 0 | output_token=iq5_ks | Mixed manual probe: token embedding and output tensors at iq5_ks.rules^output\.weight=iq5_ks ^token_embd\.weight=iq5_ks |
| 283 | c284_latectl_iq4kss |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.47 GB | 0 | late_control=iq4_kss | Mixed manual probe: late attention plus FFN gate/down control tensors at iq4_kss.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kss ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq4_kss |
| 284 | c324_tok_iq4kss |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.48 GB | 0 | token_only=iq4_kss | Mixed manual probe: token embedding tensor at iq4_kss.rules^token_embd\.weight=iq4_kss |
| 285 | c341_earlyattn_iq3kt |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.48 GB | 0 | early_attn=iq3_kt | Mixed manual probe: early full-attention routing tensors at iq3_kt.rules^blk\.(3|7|11|15)\.attn_(q|k|v|output)\.weight=iq3_kt |
| 286 | c367_fa_iq4ksr4 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.49 GB | 0 | fa=iq4_ks_r4 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_ks_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_ks_r4 |
| 287 | c347_srcsig_iq3kt |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.50 GB | 0 | src_signal=iq3_kt | Mixed manual probe: source-side RYS 15..19 linear-attention and SSM signal path at iq3_kt.rules^blk\.(15|16|17|18|19)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq3_kt |
| 288 | c730_fa_iq3kt_bndffn_q6k |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.50 GB | 0 | fa=iq3_kt, boundary_ffn=q6_K | Mixed manual probe: all full-attention Q/K/V/O tensors at iq3_kt; RYS entry, exit, and duplicate-boundary FFN tensors at q6_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq3_kt ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q6_K |
| 289 | c405_faqkv_iq4xsr8 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.50 GB | 0 | fa_qkv=iq4_xs_r8 | Mixed manual probe: full-attention Q/K/V tensors at iq4_xs_r8.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=iq4_xs_r8 |
| 290 | c401_faqkv_iq4ksr4 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.50 GB | 0 | fa_qkv=iq4_ks_r4 | Mixed manual probe: full-attention Q/K/V tensors at iq4_ks_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=iq4_ks_r4 |
| 291 | c452_latectl_iq4ksr4 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.51 GB | 0 | late_control=iq4_ks_r4 | Mixed manual probe: late attention plus FFN gate/down control tensors at iq4_ks_r4.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_ks_r4 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq4_ks_r4 |
| 292 | c179_faqk_iq4xs |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.52 GB | 0 | fa_qk=iq4_xs | Mixed manual probe: full-attention Q/K tensors at iq4_xs.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=iq4_xs |
| 293 | c362_duppath_iq4kss |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.52 GB | 0 | dup_path=iq4_kss | Mixed manual probe: duplicated-side RYS 20..24 attention, linear-attention, and SSM path at iq4_kss.rules^blk\.(20|21|22|23|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_kss |
| 294 | c249_bndffn_iq4xs |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.52 GB | 0 | boundary_ffn=iq4_xs | Mixed manual probe: RYS entry, exit, and duplicate-boundary FFN tensors at iq4_xs.rules^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=iq4_xs |
| 295 | c549_fa_iq4kss_bndffn_iq5ks |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.53 GB | 0 | fa=iq4_kss, boundary_ffn=iq5_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kss; RYS entry, exit, and duplicate-boundary FFN tensors at iq5_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kss ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=iq5_ks |
| 296 | c439_ryssig_iq4xsr8 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.53 GB | 0 | rys_signal=iq4_xs_r8 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at iq4_xs_r8.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_xs_r8 |
| 297 | c344_srcsig_iq4kss |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.53 GB | 0 | src_signal=iq4_kss | Mixed manual probe: source-side RYS 15..19 linear-attention and SSM signal path at iq4_kss.rules^blk\.(15|16|17|18|19)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_kss |
| 298 | c352_dupsig_iq4kt |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.53 GB | 0 | dup_signal=iq4_kt | Mixed manual probe: duplicated-side RYS 20..24 linear-attention and SSM signal path at iq4_kt.rules^blk\.(20|21|22|23|24)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_kt |
| 299 | c223_ryssig_iq4ks |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.53 GB | 0 | rys_signal=iq4_ks | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at iq4_ks.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_ks |
| 300 | c272_lateattn_mxfp4 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.54 GB | 0 | late_attn=mxfp4 | Mixed manual probe: late full-attention control tensors at mxfp4.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=mxfp4 |
| 301 | c263_lateattn_iq4ks |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.54 GB | 0 | late_attn=iq4_ks | Mixed manual probe: late full-attention control tensors at iq4_ks.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_ks |
| 302 | c360_srcpath_mxfp4 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.54 GB | 0 | src_path=mxfp4 | Mixed manual probe: source-side RYS 15..19 attention, linear-attention, and SSM path at mxfp4.rules^blk\.(15|16|17|18|19)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=mxfp4 |
| 303 | c355_srcpath_iq4ks |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.54 GB | 0 | src_path=iq4_ks | Mixed manual probe: source-side RYS 15..19 attention, linear-attention, and SSM path at iq4_ks.rules^blk\.(15|16|17|18|19)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_ks |
| 304 | c361_duppath_iq4ks |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.54 GB | 0 | dup_path=iq4_ks | Mixed manual probe: duplicated-side RYS 20..24 attention, linear-attention, and SSM path at iq4_ks.rules^blk\.(20|21|22|23|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_ks |
| 305 | c212_fao_mxfp4 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.54 GB | 0 | fa_o=mxfp4 | Mixed manual probe: full-attention output tensors at mxfp4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=mxfp4 |
| 306 | c342_earlyattn_mxfp4 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.54 GB | 0 | early_attn=mxfp4 | Mixed manual probe: early full-attention routing tensors at mxfp4.rules^blk\.(3|7|11|15)\.attn_(q|k|v|output)\.weight=mxfp4 |
| 307 | c199_fav_iq4xs |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.54 GB | 0 | fa_v=iq4_xs | Mixed manual probe: full-attention V tensors at iq4_xs.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=iq4_xs |
| 308 | c372_fa_iq4nlr4 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.54 GB | 0 | fa=iq4_nl_r4 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_nl_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_nl_r4 |
| 309 | c380_fa_q40r8 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.54 GB | 0 | fa=q4_0_r8 | Mixed manual probe: all full-attention Q/K/V/O tensors at q4_0_r8.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q4_0_r8 |
| 310 | c403_faqkv_iq4kr4 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.54 GB | 0 | fa_qkv=iq4_k_r4 | Mixed manual probe: full-attention Q/K/V tensors at iq4_k_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=iq4_k_r4 |
| 311 | c343_srcsig_iq4ks |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.54 GB | 0 | src_signal=iq4_ks | Mixed manual probe: source-side RYS 15..19 linear-attention and SSM signal path at iq4_ks.rules^blk\.(15|16|17|18|19)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_ks |
| 312 | c266_lateattn_iq4k |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.55 GB | 0 | late_attn=iq4_k | Mixed manual probe: late full-attention control tensors at iq4_k.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_k |
| 313 | c286_latectl_iq4k |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.55 GB | 0 | late_control=iq4_k | Mixed manual probe: late attention plus FFN gate/down control tensors at iq4_k.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_k ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq4_k |
| 314 | c420_ryspath_iq4kr4 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.55 GB | 0 | rys_path=iq4_k_r4 | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at iq4_k_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_k_r4 |
| 315 | c246_bndffn_iq4k |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.55 GB | 0 | boundary_ffn=iq4_k | Mixed manual probe: RYS entry, exit, and duplicate-boundary FFN tensors at iq4_k.rules^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=iq4_k |
| 316 | c276_lateffn_iq4k |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.55 GB | 0 | late_ffn=iq4_k | Mixed manual probe: late FFN gate/down control tensors at iq4_k.rules^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq4_k |
| 317 | c296_finalffn_iq4k |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.55 GB | 0 | final_ffn=iq4_k | Mixed manual probe: final FFN tensors at iq4_k.rules^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=iq4_k |
| 318 | c326_tok_iq4k |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.55 GB | 0 | token_only=iq4_k | Mixed manual probe: token embedding tensor at iq4_k.rules^token_embd\.weight=iq4_k |
| 319 | c389_faqk_iq4nlr4 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.55 GB | 0 | fa_qk=iq4_nl_r4 | Mixed manual probe: full-attention Q/K tensors at iq4_nl_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=iq4_nl_r4 |
| 320 | c437_ryssig_iq4kr4 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.55 GB | 0 | rys_signal=iq4_k_r4 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at iq4_k_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_k_r4 |
| 321 | c489_fav_q8kv |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.58 GB | 0 | fa_v=q8_KV | Mixed manual probe: full-attention V tensors at q8_KV.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=q8_KV |
| 322 | c351_dupsig_iq5ks |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.59 GB | 0 | dup_signal=iq5_ks | Mixed manual probe: duplicated-side RYS 20..24 linear-attention and SSM signal path at iq5_ks.rules^blk\.(20|21|22|23|24)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq5_ks |
| 323 | c265_lateattn_iq5ks |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.60 GB | 0 | late_attn=iq5_ks | Mixed manual probe: late full-attention control tensors at iq5_ks.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks |
| 324 | c337_earlyattn_iq5k |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.60 GB | 0 | early_attn=iq5_k | Mixed manual probe: early full-attention routing tensors at iq5_k.rules^blk\.(3|7|11|15)\.attn_(q|k|v|output)\.weight=iq5_k |
| 325 | c363_duppath_iq5ks |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.61 GB | 0 | dup_path=iq5_ks | Mixed manual probe: duplicated-side RYS 20..24 attention, linear-attention, and SSM path at iq5_ks.rules^blk\.(20|21|22|23|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq5_ks |
| 326 | c275_lateffn_iq5ks |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.62 GB | 0 | late_ffn=iq5_ks | Mixed manual probe: late FFN gate/down control tensors at iq5_ks.rules^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq5_ks |
| 327 | c227_ryssig_iq5k |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.64 GB | 0 | rys_signal=iq5_k | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at iq5_k.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq5_k |
| 328 | c438_ryssig_iq5kr4 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.64 GB | 0 | rys_signal=iq5_k_r4 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at iq5_k_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq5_k_r4 |
| 329 | c584_fa_iq5ks_bndffn_iq4kss |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.66 GB | 0 | fa=iq5_ks, boundary_ffn=iq4_kss | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_ks; RYS entry, exit, and duplicate-boundary FFN tensors at iq4_kss.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=iq4_kss |
| 330 | c215_ryspath_iq5ks |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.66 GB | 0 | rys_path=iq5_ks | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at iq5_ks.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq5_ks |
| 331 | c285_latectl_iq5ks |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.67 GB | 0 | late_control=iq5_ks | Mixed manual probe: late attention plus FFN gate/down control tensors at iq5_ks.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq5_ks |
| 332 | c325_tok_iq5ks |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.68 GB | 0 | token_only=iq5_ks | Mixed manual probe: token embedding tensor at iq5_ks.rules^token_embd\.weight=iq5_ks |
| 333 | c247_bndffn_iq5k |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.69 GB | 0 | boundary_ffn=iq5_k | Mixed manual probe: RYS entry, exit, and duplicate-boundary FFN tensors at iq5_k.rules^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=iq5_k |
| 334 | c268_lateattn_iq6k |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.69 GB | 0 | late_attn=iq6_k | Mixed manual probe: late full-attention control tensors at iq6_k.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq6_k |
| 335 | c177_faqk_iq5k |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.71 GB | 0 | fa_qk=iq5_k | Mixed manual probe: full-attention Q/K tensors at iq5_k.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=iq5_k |
| 336 | c396_faqk_q50r4 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.71 GB | 0 | fa_qk=q5_0_r4 | Mixed manual probe: full-attention Q/K tensors at q5_0_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=q5_0_r4 |
| 337 | c416_faqkv_q5kr4 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.71 GB | 0 | fa_qkv=q5_k_r4 | Mixed manual probe: full-attention Q/K/V tensors at q5_k_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=q5_k_r4 |
| 338 | c445_ryssig_q60 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.73 GB | 0 | rys_signal=q6_0 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at q6_0.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q6_0 |
| 339 | c446_ryssig_q60r4 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.73 GB | 0 | rys_signal=q6_0_r4 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at q6_0_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q6_0_r4 |
| 340 | c451_ryssig_q6kr4 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.73 GB | 0 | rys_signal=q6_k_r4 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at q6_k_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q6_k_r4 |
| 341 | c394_faqk_q60 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.86 GB | 0 | fa_qk=q6_0 | Mixed manual probe: full-attention Q/K tensors at q6_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=q6_0 |
| 342 | c479_outtok_q60 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.86 GB | 0 | output_token=q6_0 | Mixed manual probe: token embedding and output tensors at q6_0.rules^output\.weight=q6_0 ^token_embd\.weight=q6_0 |
| 343 | c485_outtok_q6kr4 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.88 GB | 0 | output_token=q6_k_r4 | Mixed manual probe: token embedding and output tensors at q6_k_r4.rules^output\.weight=q6_k_r4 ^token_embd\.weight=q6_k_r4 |
| 344 | c328_tok_iq6k |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.89 GB | 0 | token_only=iq6_k | Mixed manual probe: token embedding tensor at iq6_k.rules^token_embd\.weight=iq6_k |
| 345 | c188_faqkv_iq6k |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 16.89 GB | 0 | fa_qkv=iq6_k | Mixed manual probe: full-attention Q/K/V tensors at iq6_k.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=iq6_k |
| 346 | c426_ryspath_q8kr8 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 17.05 GB | 0 | rys_path=q8_k_r8 | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at q8_k_r8.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_k_r8 |
| 347 | c575_fa_iq5ks_ryssig_q8 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 17.07 GB | 0 | fa=iq5_ks, rys_signal=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_ks; RYS 15..24 linear-attention and SSM signal path at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 348 | c392_faqk_q8kr8 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 17.10 GB | 0 | fa_qk=q8_k_r8 | Mixed manual probe: full-attention Q/K tensors at q8_k_r8.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=q8_k_r8 |
| 349 | c459_latectl_q8r8 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 17.17 GB | 0 | late_control=q8_0_r8 | Mixed manual probe: late attention plus FFN gate/down control tensors at q8_0_r8.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0_r8 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0_r8 |
| 350 | c738_fa_iq3kt_lateffn_bf16 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 17.25 GB | 0 | fa=iq3_kt, late_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq3_kt; late FFN gate/down control tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq3_kt ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=bf16 |
| 351 | c63_fa_q5k_outtok_q8 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 17.72 GB | 0 | fa=q5_K, output_token=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K; token embedding and output tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K ^output\.weight=q8_0 ^token_embd\.weight=q8_0 |
| 352 | c09_union_rys_path_q8_late_control_q8 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 17.72 GB | 0 | ^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0; ^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0; ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 | Union of RYS-window path retention and late execution-control retention.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 ^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 353 | c714_fa_iq3kt_outtok_bf16 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 19.55 GB | 0 | fa=iq3_kt, output_token=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq3_kt; token embedding and output tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq3_kt ^output\.weight=bf16 ^token_embd\.weight=bf16 |
| 354 | c690_fa_iq4kt_rysffn_bf16 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 20.27 GB | 0 | fa=iq4_kt, rys_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kt; RYS 15..24 FFN tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kt ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=bf16 |
| 355 | c582_fa_iq5ks_rysffn_bf16 |
0.8443 | 0.8786 | 0.7996 | 0.8342 | 20.57 GB | 0 | fa=iq5_ks, rys_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_ks; RYS 15..24 FFN tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=bf16 |
| 356 | c757_fa_mxfp4_rysffn_iq4ks |
0.8436 | 0.8786 | 0.7980 | 0.8332 | 16.40 GB | 0 | fa=mxfp4, rys_ffn=iq4_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at mxfp4; RYS 15..24 FFN tensors at iq4_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=mxfp4 ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=iq4_ks |
| 357 | c310_outtok_iq4kt |
0.8433 | 0.8786 | 0.7974 | 0.8329 | 16.07 GB | 0 | output_token=iq4_kt | Mixed manual probe: token embedding and output tensors at iq4_kt.rules^output\.weight=iq4_kt ^token_embd\.weight=iq4_kt |
| 358 | c359_srcpath_iq3kt |
0.8433 | 0.8786 | 0.7974 | 0.8329 | 16.46 GB | 0 | src_path=iq3_kt | Mixed manual probe: source-side RYS 15..19 attention, linear-attention, and SSM path at iq3_kt.rules^blk\.(15|16|17|18|19)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq3_kt |
| 359 | c441_ryssig_iq3kr4 |
0.8433 | 0.8786 | 0.7974 | 0.8329 | 16.46 GB | 0 | rys_signal=iq3_k_r4 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at iq3_k_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq3_k_r4 |
| 360 | c256_rysffn_iq4k |
0.8433 | 0.8786 | 0.7974 | 0.8329 | 16.55 GB | 0 | rys_ffn=iq4_k | Mixed manual probe: RYS 15..24 FFN tensors at iq4_k.rules^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=iq4_k |
| 361 | c235_bndall_iq5ks |
0.8433 | 0.8786 | 0.7974 | 0.8329 | 16.69 GB | 0 | boundary_all=iq5_ks | Mixed manual probe: RYS entry, exit, and duplicate-boundary attention/SSM/FFN tensors at iq5_ks.rules^blk\.(15|19|20|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=iq5_ks |
| 362 | c589_fa_iq5ks_lateffn_iq4ks |
0.8433 | 0.8786 | 0.7974 | 0.8329 | 16.70 GB | 0 | fa=iq5_ks, late_ffn=iq4_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_ks; late FFN gate/down control tensors at iq4_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq4_ks |
| 363 | c661_fa_iq5k_lateffn_iq4ks |
0.8433 | 0.8786 | 0.7974 | 0.8329 | 16.76 GB | 0 | fa=iq5_k, late_ffn=iq4_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_k; late FFN gate/down control tensors at iq4_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_k ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq4_ks |
| 364 | c258_rysffn_iq6k |
0.8433 | 0.8786 | 0.7974 | 0.8329 | 17.27 GB | 0 | rys_ffn=iq6_k | Mixed manual probe: RYS 15..24 FFN tensors at iq6_k.rules^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=iq6_k |
| 365 | c283_latectl_iq4ks |
0.8424 | 0.8786 | 0.7952 | 0.8317 | 16.51 GB | 0 | late_control=iq4_ks | Mixed manual probe: late attention plus FFN gate/down control tensors at iq4_ks.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_ks ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq4_ks |
| 366 | c356_srcpath_iq4kss |
0.8424 | 0.8786 | 0.7952 | 0.8317 | 16.52 GB | 0 | src_path=iq4_kss | Mixed manual probe: source-side RYS 15..19 attention, linear-attention, and SSM path at iq4_kss.rules^blk\.(15|16|17|18|19)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq4_kss |
| 367 | c414_faqkv_q40r8 |
0.8424 | 0.8786 | 0.7952 | 0.8317 | 16.54 GB | 0 | fa_qkv=q4_0_r8 | Mixed manual probe: full-attention Q/K/V tensors at q4_0_r8.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=q4_0_r8 |
| 368 | c468_latectl_q6kr4 |
0.8424 | 0.8786 | 0.7952 | 0.8317 | 16.87 GB | 0 | late_control=q6_k_r4 | Mixed manual probe: late attention plus FFN gate/down control tensors at q6_k_r4.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_k_r4 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q6_k_r4 |
| 369 | c112_fav_bf16_fao_bf16 |
0.8424 | 0.8786 | 0.7952 | 0.8317 | 17.49 GB | 0 | fa_v=bf16, fa_o=bf16 | Mixed manual probe: full-attention V tensors at bf16; full-attention output tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=bf16 ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=bf16 |
| 370 | c264_lateattn_iq4kss |
0.8416 | 0.8786 | 0.7935 | 0.8307 | 16.52 GB | 0 | late_attn=iq4_kss | Mixed manual probe: late full-attention control tensors at iq4_kss.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kss |
| 371 | c69_fa_q5k_bndffn_q8 |
0.8407 | 0.8786 | 0.7915 | 0.8296 | 17.31 GB | 0 | fa=q5_K, boundary_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K; RYS entry, exit, and duplicate-boundary FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 |
| 372 | c430_ryspath_q50r4 |
0.8404 | 0.8786 | 0.7907 | 0.8292 | 16.69 GB | 0 | rys_path=q5_0_r4 | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at q5_0_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q5_0_r4 |
| 373 | c320_out_iq4kt |
0.8390 | 0.8786 | 0.7874 | 0.8273 | 16.15 GB | 0 | output_only=iq4_kt | Mixed manual probe: output tensor at iq4_kt.rules^output\.weight=iq4_kt |
| 374 | c425_ryspath_q8r8 |
0.8375 | 0.8786 | 0.7841 | 0.8254 | 17.11 GB | 0 | rys_path=q8_0_r8 | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at q8_0_r8.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0_r8 |
| 375 | c588_fa_iq5ks_bndffn_bf16 |
0.8365 | 0.9143 | 0.7352 | 0.8135 | 18.26 GB | 0 | fa=iq5_ks, boundary_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_ks; RYS entry, exit, and duplicate-boundary FFN tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=bf16 |
| 376 | c96_fa_q8_lateffn_bf16 |
0.8365 | 0.9143 | 0.7352 | 0.8135 | 18.51 GB | 0 | fa=q8_0, late_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; late FFN gate/down control tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=bf16 |
| 377 | c255_rysffn_iq5ks |
0.8363 | 0.8786 | 0.7813 | 0.8239 | 16.81 GB | 0 | rys_ffn=iq5_ks | Mixed manual probe: RYS 15..24 FFN tensors at iq5_ks.rules^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=iq5_ks |
| 378 | c455_latectl_iq5kr4 |
0.8356 | 0.8786 | 0.7796 | 0.8229 | 16.71 GB | 0 | late_control=iq5_k_r4 | Mixed manual probe: late attention plus FFN gate/down control tensors at iq5_k_r4.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_k_r4 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq5_k_r4 |
| 379 | c390_faqk_iq3kr4 |
0.8346 | 0.8786 | 0.7774 | 0.8217 | 16.39 GB | 0 | fa_qk=iq3_k_r4 | Mixed manual probe: full-attention Q/K tensors at iq3_k_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=iq3_k_r4 |
| 380 | c447_ryssig_q50r4 |
0.8344 | 0.8786 | 0.7769 | 0.8214 | 16.64 GB | 0 | rys_signal=q5_0_r4 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at q5_0_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q5_0_r4 |
| 381 | c469_outtok_iq4ksr4 |
0.8337 | 0.8429 | 0.8219 | 0.8310 | 16.15 GB | 0 | output_token=iq4_ks_r4 | Mixed manual probe: token embedding and output tensors at iq4_ks_r4.rules^output\.weight=iq4_ks_r4 ^token_embd\.weight=iq4_ks_r4 |
| 382 | c294_finalffn_iq4kss |
0.8337 | 0.8429 | 0.8219 | 0.8310 | 16.51 GB | 0 | final_ffn=iq4_kss | Mixed manual probe: final FFN tensors at iq4_kss.rules^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=iq4_kss |
| 383 | c292_latectl_mxfp4 |
0.8337 | 0.8429 | 0.8219 | 0.8310 | 16.51 GB | 0 | late_control=mxfp4 | Mixed manual probe: late attention plus FFN gate/down control tensors at mxfp4.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=mxfp4 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=mxfp4 |
| 384 | c194_fav_iq4kss |
0.8337 | 0.8429 | 0.8219 | 0.8310 | 16.54 GB | 0 | fa_v=iq4_kss | Mixed manual probe: full-attention V tensors at iq4_kss.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=iq4_kss |
| 385 | c406_faqkv_iq4nlr4 |
0.8337 | 0.8429 | 0.8219 | 0.8310 | 16.54 GB | 0 | fa_qkv=iq4_nl_r4 | Mixed manual probe: full-attention Q/K/V tensors at iq4_nl_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=iq4_nl_r4 |
| 386 | c707_fa_iq4kt_finalffn_q8 |
0.8337 | 0.8429 | 0.8219 | 0.8310 | 16.83 GB | 0 | fa=iq4_kt, final_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kt; final FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kt ^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=q8_0 |
| 387 | c462_latectl_q60 |
0.8337 | 0.8429 | 0.8219 | 0.8310 | 16.86 GB | 0 | late_control=q6_0 | Mixed manual probe: late attention plus FFN gate/down control tensors at q6_0.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q6_0 |
| 388 | c240_bndall_iq4kt |
0.8337 | 0.8786 | 0.7752 | 0.8204 | 16.46 GB | 0 | boundary_all=iq4_kt | Mixed manual probe: RYS entry, exit, and duplicate-boundary attention/SSM/FFN tensors at iq4_kt.rules^blk\.(15|19|20|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=iq4_kt |
| 389 | c432_ryspath_q4kr4 |
0.8337 | 0.8786 | 0.7752 | 0.8204 | 16.55 GB | 0 | rys_path=q4_k_r4 | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at q4_k_r4.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q4_k_r4 |
| 390 | c448_ryssig_q40r8 |
0.8337 | 0.8786 | 0.7752 | 0.8204 | 16.55 GB | 0 | rys_signal=q4_0_r8 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at q4_0_r8.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q4_0_r8 |
| 391 | c399_faqk_q5kr4 |
0.8337 | 0.8786 | 0.7752 | 0.8204 | 16.71 GB | 0 | fa_qk=q5_k_r4 | Mixed manual probe: full-attention Q/K tensors at q5_k_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=q5_k_r4 |
| 392 | c490_ryspath_q8kv |
0.8337 | 0.8786 | 0.7752 | 0.8204 | 17.04 GB | 0 | rys_path=q8_KV | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at q8_KV.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_KV |
| 393 | c410_faqkv_q8kvr8 |
0.8337 | 0.8786 | 0.7752 | 0.8204 | 17.12 GB | 0 | fa_qkv=q8_KV_r8 | Mixed manual probe: full-attention Q/K/V tensors at q8_KV_r8.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=q8_KV_r8 |
| 394 | c126_bndall_bf16_lateffn_q8 |
0.8337 | 0.8786 | 0.7752 | 0.8204 | 19.05 GB | 0 | boundary_all=bf16, late_ffn=q8_0 | Mixed manual probe: RYS entry, exit, and duplicate-boundary attention/SSM/FFN tensors at bf16; late FFN gate/down control tensors at q8_0.rules^blk\.(15|19|20|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=bf16 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 395 | c257_rysffn_iq5k |
0.8336 | 0.8614 | 0.7974 | 0.8254 | 16.89 GB | 0 | rys_ffn=iq5_k | Mixed manual probe: RYS 15..24 FFN tensors at iq5_k.rules^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=iq5_k |
| 396 | c710_fa_iq3kt_outtok_iq4kss |
0.8322 | 0.8571 | 0.7996 | 0.8248 | 15.74 GB | 0 | fa=iq3_kt, output_token=iq4_kss | Mixed manual probe: all full-attention Q/K/V/O tensors at iq3_kt; token embedding and output tensors at iq4_kss.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq3_kt ^output\.weight=iq4_kss ^token_embd\.weight=iq4_kss |
| 397 | c400_faqk_q6kr4 |
0.8298 | 0.8786 | 0.7663 | 0.8154 | 16.87 GB | 0 | fa_qk=q6_k_r4 | Mixed manual probe: full-attention Q/K tensors at q6_k_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=q6_k_r4 |
| 398 | c261_rysffn_iq3kt |
0.8281 | 0.8500 | 0.7996 | 0.8217 | 16.10 GB | 0 | rys_ffn=iq3_kt | Mixed manual probe: RYS 15..24 FFN tensors at iq3_kt.rules^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=iq3_kt |
| 399 | c388_faqk_iq4xsr8 |
0.8281 | 0.8500 | 0.7996 | 0.8217 | 16.52 GB | 0 | fa_qk=iq4_xs_r8 | Mixed manual probe: full-attention Q/K tensors at iq4_xs_r8.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=iq4_xs_r8 |
| 400 | c202_fav_mxfp4 |
0.8276 | 0.8429 | 0.8078 | 0.8231 | 16.54 GB | 0 | fa_v=mxfp4 | Mixed manual probe: full-attention V tensors at mxfp4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=mxfp4 |
| 401 | c83_fa_q6k_lateffn_q8 |
0.8268 | 0.9143 | 0.7130 | 0.8010 | 17.39 GB | 0 | fa=q6_K, late_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K; late FFN gate/down control tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 402 | c293_finalffn_iq4ks |
0.8241 | 0.8429 | 0.7996 | 0.8185 | 16.53 GB | 0 | final_ffn=iq4_ks | Mixed manual probe: final FFN tensors at iq4_ks.rules^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=iq4_ks |
| 403 | c694_fa_iq4kt_bndffn_q6k |
0.8241 | 0.8429 | 0.7996 | 0.8185 | 16.70 GB | 0 | fa=iq4_kt, boundary_ffn=q6_K | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kt; RYS entry, exit, and duplicate-boundary FFN tensors at q6_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kt ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q6_K |
| 404 | c745_fa_mxfp4_outtok_iq4ks |
0.8234 | 0.8429 | 0.7980 | 0.8176 | 16.08 GB | 0 | fa=mxfp4, output_token=iq4_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at mxfp4; token embedding and output tensors at iq4_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=mxfp4 ^output\.weight=iq4_ks ^token_embd\.weight=iq4_ks |
| 405 | c517_fa_iq4ks_lateffn_iq4ks |
0.8163 | 0.8786 | 0.7352 | 0.7979 | 16.46 GB | 0 | fa=iq4_ks, late_ffn=iq4_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_ks; late FFN gate/down control tensors at iq4_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_ks ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq4_ks |
| 406 | c163_fa_iq4ks |
0.8163 | 0.8786 | 0.7352 | 0.7979 | 16.49 GB | 0 | fa=iq4_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_ks |
| 407 | c186_faqkv_iq4k |
0.8163 | 0.8786 | 0.7352 | 0.7979 | 16.54 GB | 0 | fa_qkv=iq4_k | Mixed manual probe: full-attention Q/K/V tensors at iq4_k.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=iq4_k |
| 408 | c454_latectl_iq4kr4 |
0.8163 | 0.8786 | 0.7352 | 0.7979 | 16.55 GB | 0 | late_control=iq4_k_r4 | Mixed manual probe: late attention plus FFN gate/down control tensors at iq4_k_r4.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_k_r4 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq4_k_r4 |
| 409 | c457_latectl_iq4nlr4 |
0.8163 | 0.8786 | 0.7352 | 0.7979 | 16.55 GB | 0 | late_control=iq4_nl_r4 | Mixed manual probe: late attention plus FFN gate/down control tensors at iq4_nl_r4.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_nl_r4 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq4_nl_r4 |
| 410 | c297_finalffn_iq5k |
0.8163 | 0.8786 | 0.7352 | 0.7979 | 16.66 GB | 0 | final_ffn=iq5_k | Mixed manual probe: final FFN tensors at iq5_k.rules^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=iq5_k |
| 411 | c175_faqk_iq5ks |
0.8163 | 0.8786 | 0.7352 | 0.7979 | 16.67 GB | 0 | fa_qk=iq5_ks | Mixed manual probe: full-attention Q/K tensors at iq5_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=iq5_ks |
| 412 | c217_ryspath_iq5k |
0.8163 | 0.8786 | 0.7352 | 0.7979 | 16.69 GB | 0 | rys_path=iq5_k | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at iq5_k.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq5_k |
| 413 | c585_fa_iq5ks_bndffn_iq5ks |
0.8163 | 0.8786 | 0.7352 | 0.7979 | 16.82 GB | 0 | fa=iq5_ks, boundary_ffn=iq5_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_ks; RYS entry, exit, and duplicate-boundary FFN tensors at iq5_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=iq5_ks |
| 414 | c586_fa_iq5ks_bndffn_q6k |
0.8163 | 0.8786 | 0.7352 | 0.7979 | 17.00 GB | 0 | fa=iq5_ks, boundary_ffn=q6_K | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_ks; RYS entry, exit, and duplicate-boundary FFN tensors at q6_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q6_K |
| 415 | c168_fa_iq6k |
0.8163 | 0.8786 | 0.7352 | 0.7979 | 17.04 GB | 0 | fa=iq6_k | Mixed manual probe: all full-attention Q/K/V/O tensors at iq6_k.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq6_k |
| 416 | c684_fa_iq4kt_ryssig_bf16 |
0.8163 | 0.8786 | 0.7352 | 0.7979 | 17.43 GB | 0 | fa=iq4_kt, rys_signal=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kt; RYS 15..24 linear-attention and SSM signal path at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kt ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 |
| 417 | c374_fa_q8r8 |
0.8163 | 0.8786 | 0.7352 | 0.7979 | 17.49 GB | 0 | fa=q8_0_r8 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0_r8.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0_r8 |
| 418 | c522_fa_iq4ks_lateffn_bf16 |
0.8163 | 0.8786 | 0.7352 | 0.7979 | 17.51 GB | 0 | fa=iq4_ks, late_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_ks; late FFN gate/down control tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_ks ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=bf16 |
| 419 | c594_fa_iq5ks_lateffn_bf16 |
0.8163 | 0.8786 | 0.7352 | 0.7979 | 17.75 GB | 0 | fa=iq5_ks, late_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_ks; late FFN gate/down control tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=bf16 |
| 420 | c371_fa_iq4xsr8 |
0.8135 | 0.8429 | 0.7752 | 0.8048 | 16.48 GB | 0 | fa=iq4_xs_r8 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_xs_r8.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_xs_r8 |
| 421 | c245_bndffn_iq5ks |
0.8135 | 0.8429 | 0.7752 | 0.8048 | 16.66 GB | 0 | boundary_ffn=iq5_ks | Mixed manual probe: RYS entry, exit, and duplicate-boundary FFN tensors at iq5_ks.rules^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=iq5_ks |
| 422 | c165_fa_iq5ks |
0.8086 | 0.8786 | 0.7174 | 0.7879 | 16.72 GB | 0 | fa=iq5_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks |
| 423 | c304_outtok_iq4kss |
0.8066 | 0.8786 | 0.7130 | 0.7854 | 16.07 GB | 0 | output_token=iq4_kss | Mixed manual probe: token embedding and output tensors at iq4_kss.rules^output\.weight=iq4_kss ^token_embd\.weight=iq4_kss |
| 424 | c415_faqkv_q4kr4 |
0.8066 | 0.8786 | 0.7130 | 0.7854 | 16.54 GB | 0 | fa_qkv=q4_k_r4 | Mixed manual probe: full-attention Q/K/V tensors at q4_k_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=q4_k_r4 |
| 425 | c357_srcpath_iq5ks |
0.8066 | 0.8786 | 0.7130 | 0.7854 | 16.61 GB | 0 | src_path=iq5_ks | Mixed manual probe: source-side RYS 15..19 attention, linear-attention, and SSM path at iq5_ks.rules^blk\.(15|16|17|18|19)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=iq5_ks |
| 426 | c185_faqkv_iq5ks |
0.8066 | 0.8786 | 0.7130 | 0.7854 | 16.67 GB | 0 | fa_qkv=iq5_ks | Mixed manual probe: full-attention Q/K/V tensors at iq5_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=iq5_ks |
| 427 | c402_faqkv_iq5ksr4 |
0.8066 | 0.8786 | 0.7130 | 0.7854 | 16.67 GB | 0 | fa_qkv=iq5_ks_r4 | Mixed manual probe: full-attention Q/K/V tensors at iq5_ks_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=iq5_ks_r4 |
| 428 | c726_fa_iq3kt_rysffn_bf16 |
0.8066 | 0.8786 | 0.7130 | 0.7854 | 20.06 GB | 0 | fa=iq3_kt, rys_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq3_kt; RYS 15..24 FFN tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq3_kt ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=bf16 |
| 429 | c314_out_iq4kss |
0.7961 | 0.8429 | 0.7352 | 0.7823 | 16.15 GB | 0 | output_only=iq4_kss | Mixed manual probe: output tensor at iq4_kss.rules^output\.weight=iq4_kss |
| 430 | c458_latectl_iq3kr4 |
0.7960 | 0.8786 | 0.6885 | 0.7717 | 16.39 GB | 0 | late_control=iq3_k_r4 | Mixed manual probe: late attention plus FFN gate/down control tensors at iq3_k_r4.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq3_k_r4 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=iq3_k_r4 |
| 431 | c487_faqk_q8kv |
0.7960 | 0.8786 | 0.6885 | 0.7717 | 17.09 GB | 0 | fa_qk=q8_KV | Mixed manual probe: full-attention Q/K tensors at q8_KV.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=q8_KV |
| 432 | c222_ryspath_mxfp4 |
0.7858 | 0.8571 | 0.6930 | 0.7648 | 16.52 GB | 0 | rys_path=mxfp4 | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at mxfp4.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=mxfp4 |
| 433 | c376_fa_q8kvr8 |
0.7451 | 0.8786 | 0.6863 | 0.7704 | 17.37 GB | 1 | fa=q8_KV_r8 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_KV_r8.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_KV_r8 |
| 434 | c378_fa_q60r4 |
0.6776 | 0.9143 | 0.5996 | 0.7373 | 17.01 GB | 2 | fa=q6_0_r4 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_0_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_0_r4 |
| 435 | c383_fa_q6kr4 |
0.6776 | 0.9143 | 0.5996 | 0.7373 | 17.03 GB | 2 | fa=q6_k_r4 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_k_r4.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_k_r4 |
| 436 | c486_fa_q8kv |
0.5301 | 0.7357 | 0.6078 | 0.6637 | 17.37 GB | 3 | fa=q8_KV | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_KV.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_KV |
| 437 | c375_fa_q8kr8 |
0.5266 | 0.7357 | 0.5996 | 0.6592 | 17.38 GB | 3 | fa=q8_k_r8 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_k_r8.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_k_r8 |
| 438 | c687_fa_iq4kt_rysffn_iq5ks |
0.3271 | 0.6286 | 0.5100 | 0.5619 | 16.68 GB | 5 | fa=iq4_kt, rys_ffn=iq5_ks | Mixed manual probe: all full-attention Q/K/V/O tensors at iq4_kt; RYS 15..24 FFN tensors at iq5_ks.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq4_kt ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=iq5_ks |
| 439 | c488_faqkv_q8kv |
0.0000 | 0.0000 | 0.3774 | 0.2123 | 17.12 GB | 11 | fa_qkv=q8_KV | Mixed manual probe: full-attention Q/K/V tensors at q8_KV.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=q8_KV |
| 440 | c581_fa_iq5ks_rysffn_q8 |
0.0000 | 0.1286 | 0.4167 | 0.2906 | 18.06 GB | 11 | fa=iq5_ks, rys_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at iq5_ks; RYS 15..24 FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=iq5_ks ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q8_0 |
Stable promotion-screen table
This table contains all 163 rows in the stable promotion summary used for the 163-run / 1,956-prompt accounting.
| Rank | Candidate | Weighted | Habit | Hard | Mean | Size | Errors | Recipe spec | Description / exact rules |
|---|---|---|---|---|---|---|---|---|---|
| 1 | c76_fa_q6k_outtok_bf16 |
0.8413 | 0.9167 | 0.7394 | 0.8281 | 20.36 GB | 0 | fa=q6_K, output_token=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K; token embedding and output tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^output\.weight=bf16 ^token_embd\.weight=bf16 |
| 2 | c21_fao_q6k |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 16.70 GB | 0 | fa_o=q6_K | Mixed manual probe: full-attention output tensors at q6_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=q6_K |
| 3 | c22_fao_q8 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 16.84 GB | 0 | fa_o=q8_0 | Mixed manual probe: full-attention output tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=q8_0 |
| 4 | c69_fa_q5k_bndffn_q8 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 17.31 GB | 0 | fa=q5_K, boundary_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K; RYS entry, exit, and duplicate-boundary FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 |
| 5 | c23_fao_bf16 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 17.37 GB | 0 | fa_o=bf16 | Mixed manual probe: full-attention output tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=bf16 |
| 6 | c112_fav_bf16_fao_bf16 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 17.49 GB | 0 | fa_v=bf16, fa_o=bf16 | Mixed manual probe: full-attention V tensors at bf16; full-attention output tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=bf16 ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=bf16 |
| 7 | c89_fa_q8_ryssig_q8 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 17.83 GB | 0 | fa=q8_0, rys_signal=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; RYS 15..24 linear-attention and SSM signal path at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 8 | c114_fao_bf16_faqk_q8 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 17.98 GB | 0 | fa_o=bf16, fa_qk=q8_0 | Mixed manual probe: full-attention output tensors at bf16; full-attention Q/K tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=bf16 ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=q8_0 |
| 9 | c93_fa_q8_bndffn_q8 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 18.02 GB | 0 | fa=q8_0, boundary_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; RYS entry, exit, and duplicate-boundary FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 |
| 10 | c152_fav_bf16_fao_bf16_bndffn_q8 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 18.03 GB | 0 | fa_v=bf16, fa_o=bf16, boundary_ffn=q8_0 | Mixed manual probe: full-attention V tensors at bf16; full-attention output tensors at bf16; RYS entry, exit, and duplicate-boundary FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=bf16 ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=bf16 ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 |
| 11 | c154_fao_bf16_faqk_q8_lateffn_q8 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 18.34 GB | 0 | fa_o=bf16, fa_qk=q8_0, late_ffn=q8_0 | Mixed manual probe: full-attention output tensors at bf16; full-attention Q/K tensors at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=bf16 ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 12 | c87_fa_q8_outtok_q8 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 18.43 GB | 0 | fa=q8_0, output_token=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; token embedding and output tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 |
| 13 | c96_fa_q8_lateffn_bf16 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 18.51 GB | 0 | fa=q8_0, late_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; late FFN gate/down control tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=bf16 |
| 14 | c142_fa_q8_outtok_q8_lateffn_q8 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 18.79 GB | 0 | fa=q8_0, output_token=q8_0, late_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; token embedding and output tensors at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 15 | c126_bndall_bf16_lateffn_q8 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 19.05 GB | 0 | boundary_all=bf16, late_ffn=q8_0 | Mixed manual probe: RYS entry, exit, and duplicate-boundary attention/SSM/FFN tensors at bf16; late FFN gate/down control tensors at q8_0.rules^blk\.(15|19|20|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=bf16 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 16 | c05_all_fullattn_bf16 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 19.26 GB | 0 | ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 | Full-attention ceiling probe: all full-attention Q/K/V/O tensors remain BF16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 |
| 17 | c101_fa_bf16_ryssig_q8 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 19.60 GB | 0 | fa=bf16, rys_signal=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; RYS 15..24 linear-attention and SSM signal path at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 18 | c56_outtok_bf16 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 19.88 GB | 0 | output_token=bf16 | Mixed manual probe: token embedding and output tensors at bf16.rules^output\.weight=bf16 ^token_embd\.weight=bf16 |
| 19 | c64_fa_q5k_outtok_bf16 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 20.11 GB | 0 | fa=q5_K, output_token=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K; token embedding and output tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K ^output\.weight=bf16 ^token_embd\.weight=bf16 |
| 20 | c99_fa_bf16_outtok_q8 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 20.20 GB | 0 | fa=bf16, output_token=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; token embedding and output tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 |
| 21 | c102_fa_bf16_ryssig_bf16 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 20.26 GB | 0 | fa=bf16, rys_signal=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; RYS 15..24 linear-attention and SSM signal path at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 |
| 22 | c148_fa_bf16_outtok_q8_lateffn_q8 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 20.56 GB | 0 | fa=bf16, output_token=q8_0, late_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; token embedding and output tensors at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 23 | c147_fa_bf16_outtok_q8_bndffn_q8 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 20.73 GB | 0 | fa=bf16, output_token=q8_0, boundary_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; token embedding and output tensors at q8_0; RYS entry, exit, and duplicate-boundary FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 |
| 24 | c88_fa_q8_outtok_bf16 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 20.81 GB | 0 | fa=q8_0, output_token=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; token embedding and output tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^output\.weight=bf16 ^token_embd\.weight=bf16 |
| 25 | c100_fa_bf16_outtok_bf16 |
0.8385 | 0.9167 | 0.7328 | 0.8247 | 22.58 GB | 0 | fa=bf16, output_token=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; token embedding and output tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^output\.weight=bf16 ^token_embd\.weight=bf16 |
| 26 | c06_output_token_q8_fullattn_q8 |
0.8357 | 0.9167 | 0.7261 | 0.8214 | 18.43 GB | 0 | ^output\.weight=q8_0; ^token_embd\.weight=q8_0; ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 | Protect expression/input surfaces plus full-attention routing: token/output Q8_0 and full-attention Q/K/V/O Q8_0.rules^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 |
| 27 | c106_fa_bf16_bndffn_bf16 |
0.8333 | 0.9167 | 0.7206 | 0.8186 | 20.79 GB | 0 | fa=bf16, boundary_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; RYS entry, exit, and duplicate-boundary FFN tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=bf16 |
| 28 | c35_bndall_q6k |
0.8243 | 0.9167 | 0.6994 | 0.8081 | 16.94 GB | 0 | boundary_all=q6_K | Mixed manual probe: RYS entry, exit, and duplicate-boundary attention/SSM/FFN tensors at q6_K.rules^blk\.(15|19|20|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=q6_K |
| 29 | c111_fav_bf16_fao_q8 |
0.8243 | 0.9167 | 0.6994 | 0.8081 | 16.96 GB | 0 | fa_v=bf16, fa_o=q8_0 | Mixed manual probe: full-attention V tensors at bf16; full-attention output tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=bf16 ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=q8_0 |
| 30 | c03_habit_boundary_q8 |
0.8243 | 0.9167 | 0.6994 | 0.8081 | 17.30 GB | 0 | ^blk\.(15|19|20|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=q8_0 | Protect the RYS entry/exit boundaries and duplicated boundary layers with Q8_0 across attention, SSM, and FFN tensors.rules^blk\.(15|19|20|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=q8_0 |
| 31 | c45_lateattn_bf16 |
0.8243 | 0.9167 | 0.6994 | 0.8081 | 17.31 GB | 0 | late_attn=bf16 | Mixed manual probe: late full-attention control tensors at bf16.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 |
| 32 | c77_fa_q6k_ryssig_q8 |
0.8243 | 0.9167 | 0.6994 | 0.8081 | 17.38 GB | 0 | fa=q6_K, rys_signal=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K; RYS 15..24 linear-attention and SSM signal path at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 33 | c83_fa_q6k_lateffn_q8 |
0.8243 | 0.9167 | 0.6994 | 0.8081 | 17.39 GB | 0 | fa=q6_K, late_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K; late FFN gate/down control tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 34 | c63_fa_q5k_outtok_q8 |
0.8243 | 0.9167 | 0.6994 | 0.8081 | 17.72 GB | 0 | fa=q5_K, output_token=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K; token embedding and output tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K ^output\.weight=q8_0 ^token_embd\.weight=q8_0 |
| 35 | c15_faqk_bf16 |
0.8243 | 0.9167 | 0.6994 | 0.8081 | 18.32 GB | 0 | fa_qk=bf16 | Mixed manual probe: full-attention Q/K tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=bf16 |
| 36 | c151_fav_bf16_fao_bf16_outtok_q8 |
0.8243 | 0.9167 | 0.6994 | 0.8081 | 18.44 GB | 0 | fa_v=bf16, fa_o=bf16, output_token=q8_0 | Mixed manual probe: full-attention V tensors at bf16; full-attention output tensors at bf16; token embedding and output tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=bf16 ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_output\.weight=bf16 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 |
| 37 | c141_fa_q8_outtok_q8_bndffn_q8 |
0.8243 | 0.9167 | 0.6994 | 0.8081 | 18.97 GB | 0 | fa=q8_0, output_token=q8_0, boundary_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; token embedding and output tensors at q8_0; RYS entry, exit, and duplicate-boundary FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 |
| 38 | c105_fa_bf16_bndffn_q8 |
0.8243 | 0.9167 | 0.6994 | 0.8081 | 19.79 GB | 0 | fa=bf16, boundary_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; RYS entry, exit, and duplicate-boundary FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 |
| 39 | c145_fa_bf16_outtok_q8_ryssig_q8 |
0.8243 | 0.9167 | 0.6994 | 0.8081 | 20.55 GB | 0 | fa=bf16, output_token=q8_0, rys_signal=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; token embedding and output tensors at q8_0; RYS 15..24 linear-attention and SSM signal path at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 40 | c01_rys_window_path_q8 |
0.8174 | 0.8750 | 0.7394 | 0.8072 | 17.11 GB | 0 | ^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 | Protect the original plus duplicated RYS 15..24 path on attention, linear-attention, and SSM tensors with Q8_0.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 41 | c09_union_rys_path_q8_late_control_q8 |
0.8174 | 0.8750 | 0.7394 | 0.8072 | 17.72 GB | 0 | ^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0; ^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0; ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 | Union of RYS-window path retention and late execution-control retention.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 ^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 42 | c18_fav_q6k |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 16.57 GB | 0 | fa_v=q6_K | Mixed manual probe: full-attention V tensors at q6_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=q6_K |
| 43 | c20_fav_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 16.68 GB | 0 | fa_v=bf16 | Mixed manual probe: full-attention V tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=bf16 |
| 44 | c27_srcsig_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 16.73 GB | 0 | src_signal=q8_0 | Mixed manual probe: source-side RYS 15..19 linear-attention and SSM signal path at q8_0.rules^blk\.(15|16|17|18|19)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 45 | c46_lateffn_q6k |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 16.74 GB | 0 | late_ffn=q6_K | Mixed manual probe: late FFN gate/down control tensors at q6_K.rules^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q6_K |
| 46 | c61_earlyattn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 16.76 GB | 0 | early_attn=q8_0 | Mixed manual probe: early full-attention routing tensors at q8_0.rules^blk\.(3|7|11|15)\.attn_(q|k|v|output)\.weight=q8_0 |
| 47 | c37_bndffn_q6k |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 16.83 GB | 0 | boundary_ffn=q6_K | Mixed manual probe: RYS entry, exit, and duplicate-boundary FFN tensors at q6_K.rules^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q6_K |
| 48 | c31_srcpath_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 16.83 GB | 0 | src_path=q8_0 | Mixed manual probe: source-side RYS 15..19 attention, linear-attention, and SSM path at q8_0.rules^blk\.(15|16|17|18|19)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 49 | c33_duppath_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 16.83 GB | 0 | dup_path=q8_0 | Mixed manual probe: duplicated-side RYS 20..24 attention, linear-attention, and SSM path at q8_0.rules^blk\.(20|21|22|23|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 50 | c57_out_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 16.86 GB | 0 | output_only=q8_0 | Mixed manual probe: output tensor at q8_0.rules^output\.weight=q8_0 |
| 51 | c49_latectl_q6k |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 16.87 GB | 0 | late_control=q6_K | Mixed manual probe: late attention plus FFN gate/down control tensors at q6_K.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q6_K |
| 52 | c13_faqk_q6k |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 16.87 GB | 0 | fa_qk=q6_K | Mixed manual probe: full-attention Q/K tensors at q6_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=q6_K |
| 53 | c54_outtok_q6k |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 16.88 GB | 0 | output_token=q6_K | Mixed manual probe: token embedding and output tensors at q6_K.rules^output\.weight=q6_K ^token_embd\.weight=q6_K |
| 54 | c25_ryssig_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 16.90 GB | 0 | rys_signal=q8_0 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at q8_0.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 55 | c47_lateffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 16.91 GB | 0 | late_ffn=q8_0 | Mixed manual probe: late FFN gate/down control tensors at q8_0.rules^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 56 | c52_finalffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 16.96 GB | 0 | final_ffn=q8_0 | Mixed manual probe: final FFN tensors at q8_0.rules^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=q8_0 |
| 57 | c12_fa_q6k |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.03 GB | 0 | fa=q6_K | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K |
| 58 | c28_srcsig_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.05 GB | 0 | src_signal=bf16 | Mixed manual probe: source-side RYS 15..19 linear-attention and SSM signal path at bf16.rules^blk\.(15|16|17|18|19)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 |
| 59 | c65_fa_q5k_ryssig_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.13 GB | 0 | fa=q5_K, rys_signal=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K; RYS 15..24 linear-attention and SSM signal path at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 60 | c62_earlyattn_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.16 GB | 0 | early_attn=bf16 | Mixed manual probe: early full-attention routing tensors at bf16.rules^blk\.(3|7|11|15)\.attn_(q|k|v|output)\.weight=bf16 |
| 61 | c14_faqk_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.17 GB | 0 | fa_qk=q8_0 | Mixed manual probe: full-attention Q/K tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=q8_0 |
| 62 | c73_fa_q5k_finalffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.18 GB | 0 | fa=q5_K, final_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K; final FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K ^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=q8_0 |
| 63 | c16_faqkv_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.20 GB | 0 | fa_qkv=q8_0 | Mixed manual probe: full-attention Q/K/V tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=q8_0 |
| 64 | c116_srcsig_q8_dupsig_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.23 GB | 0 | src_signal=q8_0, dup_signal=bf16 | Mixed manual probe: source-side RYS 15..19 linear-attention and SSM signal path at q8_0; duplicated-side RYS 20..24 linear-attention and SSM signal path at bf16.rules^blk\.(15|16|17|18|19)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 ^blk\.(20|21|22|23|24)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 |
| 65 | c113_fav_bf16_faqk_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.29 GB | 0 | fa_v=bf16, fa_qk=q8_0 | Mixed manual probe: full-attention V tensors at bf16; full-attention Q/K tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=bf16 ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=q8_0 |
| 66 | c32_srcpath_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.35 GB | 0 | src_path=bf16 | Mixed manual probe: source-side RYS 15..19 attention, linear-attention, and SSM path at bf16.rules^blk\.(15|16|17|18|19)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 |
| 67 | c85_fa_q6k_finalffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.43 GB | 0 | fa=q6_K, final_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K; final FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=q8_0 |
| 68 | c67_fa_q5k_rysffn_q6k |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.47 GB | 0 | fa=q5_K, rys_ffn=q6_K | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K; RYS 15..24 FFN tensors at q6_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q6_K |
| 69 | c131_fa_q5k_ryssig_q8_lateffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.48 GB | 0 | fa=q5_K, rys_signal=q8_0, late_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K; RYS 15..24 linear-attention and SSM signal path at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 70 | c02_all_fullattn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.49 GB | 0 | ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 | Protect all full-attention Q/K/V/O tensors with Q8_0, leaving recurrent and FFN tensors at IQ4_NL mix defaults.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 |
| 71 | c55_outtok_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.50 GB | 0 | output_token=q8_0 | Mixed manual probe: token embedding and output tensors at q8_0.rules^output\.weight=q8_0 ^token_embd\.weight=q8_0 |
| 72 | c26_ryssig_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.55 GB | 0 | rys_signal=bf16 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at bf16.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 |
| 73 | c81_fa_q6k_bndffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.56 GB | 0 | fa=q6_K, boundary_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K; RYS entry, exit, and duplicate-boundary FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 |
| 74 | c48_lateffn_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.58 GB | 0 | late_ffn=bf16 | Mixed manual probe: late FFN gate/down control tensors at bf16.rules^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=bf16 |
| 75 | c156_srcsig_bf16_dupsig_q8_lateffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.58 GB | 0 | src_signal=bf16, dup_signal=q8_0, late_ffn=q8_0 | Mixed manual probe: source-side RYS 15..19 linear-attention and SSM signal path at bf16; duplicated-side RYS 20..24 linear-attention and SSM signal path at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(15|16|17|18|19)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 ^blk\.(20|21|22|23|24)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 76 | c117_srcpath_bf16_duppath_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.63 GB | 0 | src_path=bf16, dup_path=q8_0 | Mixed manual probe: source-side RYS 15..19 attention, linear-attention, and SSM path at bf16; duplicated-side RYS 20..24 attention, linear-attention, and SSM path at q8_0.rules^blk\.(15|16|17|18|19)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 ^blk\.(20|21|22|23|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 77 | c118_srcpath_q8_duppath_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.63 GB | 0 | src_path=q8_0, dup_path=bf16 | Mixed manual probe: source-side RYS 15..19 attention, linear-attention, and SSM path at q8_0; duplicated-side RYS 20..24 attention, linear-attention, and SSM path at bf16.rules^blk\.(15|16|17|18|19)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 ^blk\.(20|21|22|23|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 |
| 78 | c125_bndall_q8_lateffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.65 GB | 0 | boundary_all=q8_0, late_ffn=q8_0 | Mixed manual probe: RYS entry, exit, and duplicate-boundary attention/SSM/FFN tensors at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(15|19|20|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 79 | c121_lateattn_bf16_lateffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.66 GB | 0 | late_attn=bf16, late_ffn=q8_0 | Mixed manual probe: late full-attention control tensors at bf16; late FFN gate/down control tensors at q8_0.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 80 | c53_finalffn_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.71 GB | 0 | final_ffn=bf16 | Mixed manual probe: final FFN tensors at bf16.rules^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=bf16 |
| 81 | c79_fa_q6k_rysffn_q6k |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.72 GB | 0 | fa=q6_K, rys_ffn=q6_K | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K; RYS 15..24 FFN tensors at q6_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q6_K |
| 82 | c137_fa_q6k_ryssig_q8_lateffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.73 GB | 0 | fa=q6_K, rys_signal=q8_0, late_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K; RYS 15..24 linear-attention and SSM signal path at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 83 | c72_fa_q5k_lateffn_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.80 GB | 0 | fa=q5_K, late_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K; late FFN gate/down control tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=bf16 |
| 84 | c122_lateattn_q8_lateffn_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.84 GB | 0 | late_attn=q8_0, late_ffn=bf16 | Mixed manual probe: late full-attention control tensors at q8_0; late FFN gate/down control tensors at bf16.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=bf16 |
| 85 | c95_fa_q8_lateffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.84 GB | 0 | fa=q8_0, late_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 86 | c97_fa_q8_finalffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.89 GB | 0 | fa=q8_0, final_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; final FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=q8_0 |
| 87 | c41_rysffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.89 GB | 0 | rys_ffn=q8_0 | Mixed manual probe: RYS 15..24 FFN tensors at q8_0.rules^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q8_0 |
| 88 | c138_fa_q6k_bndffn_q8_lateffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.92 GB | 0 | fa=q6_K, boundary_ffn=q8_0, late_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K; RYS entry, exit, and duplicate-boundary FFN tensors at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 89 | c74_fa_q5k_finalffn_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.93 GB | 0 | fa=q5_K, final_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K; final FFN tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K ^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=bf16 |
| 90 | c75_fa_q6k_outtok_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 17.97 GB | 0 | fa=q6_K, output_token=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K; token embedding and output tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^output\.weight=q8_0 ^token_embd\.weight=q8_0 |
| 91 | c78_fa_q6k_ryssig_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.03 GB | 0 | fa=q6_K, rys_signal=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K; RYS 15..24 linear-attention and SSM signal path at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 |
| 92 | c84_fa_q6k_lateffn_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.05 GB | 0 | fa=q6_K, late_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K; late FFN gate/down control tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=bf16 |
| 93 | c58_out_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.05 GB | 0 | output_only=bf16 | Mixed manual probe: output tensor at bf16.rules^output\.weight=bf16 |
| 94 | c127_fa_q5k_outtok_q8_ryssig_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.07 GB | 0 | fa=q5_K, output_token=q8_0, rys_signal=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K; token embedding and output tensors at q8_0; RYS 15..24 linear-attention and SSM signal path at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 95 | c39_bndffn_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.09 GB | 0 | boundary_ffn=bf16 | Mixed manual probe: RYS entry, exit, and duplicate-boundary FFN tensors at bf16.rules^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=bf16 |
| 96 | c123_latectl_q8_outtok_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.11 GB | 0 | late_control=q8_0, output_token=q8_0 | Mixed manual probe: late attention plus FFN gate/down control tensors at q8_0; token embedding and output tensors at q8_0.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 |
| 97 | c68_fa_q5k_rysffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.12 GB | 0 | fa=q5_K, rys_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K; RYS 15..24 FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q8_0 |
| 98 | c04_rys_window_path_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.15 GB | 0 | ^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 | Upper probe for habit retention: keep the RYS 15..24 attention/linear/SSM path at BF16.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 |
| 99 | c91_fa_q8_rysffn_q6k |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.18 GB | 0 | fa=q8_0, rys_ffn=q6_K | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; RYS 15..24 FFN tensors at q6_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q6_K |
| 100 | c143_fa_q8_ryssig_q8_lateffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.19 GB | 0 | fa=q8_0, rys_signal=q8_0, late_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; RYS 15..24 linear-attention and SSM signal path at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 101 | c153_fav_bf16_faqk_q8_outtok_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.24 GB | 0 | fa_v=bf16, fa_qk=q8_0, output_token=q8_0 | Mixed manual probe: full-attention V tensors at bf16; full-attention Q/K tensors at q8_0; token embedding and output tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=bf16 ^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k)\.weight=q8_0 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 |
| 102 | c136_fa_q6k_outtok_q8_lateffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.33 GB | 0 | fa=q6_K, output_token=q8_0, late_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K; token embedding and output tensors at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 103 | c144_fa_q8_bndffn_q8_lateffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.38 GB | 0 | fa=q8_0, boundary_ffn=q8_0, late_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; RYS entry, exit, and duplicate-boundary FFN tensors at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 104 | c17_faqkv_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.44 GB | 0 | fa_qkv=bf16 | Mixed manual probe: full-attention Q/K/V tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v)\.weight=bf16 |
| 105 | c07_rys_window_all_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.45 GB | 0 | ^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=q8_0 | Broad local recovery: the full RYS 15..24 window gets Q8_0 on attention, SSM, and FFN tensors.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=q8_0 |
| 106 | c90_fa_q8_ryssig_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.49 GB | 0 | fa=q8_0, rys_signal=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; RYS 15..24 linear-attention and SSM signal path at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 |
| 107 | c135_fa_q6k_outtok_q8_bndffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.51 GB | 0 | fa=q6_K, output_token=q8_0, boundary_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K; token embedding and output tensors at q8_0; RYS entry, exit, and duplicate-boundary FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 |
| 108 | c82_fa_q6k_bndffn_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.57 GB | 0 | fa=q6_K, boundary_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K; RYS entry, exit, and duplicate-boundary FFN tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=bf16 |
| 109 | c98_fa_q8_finalffn_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.64 GB | 0 | fa=q8_0, final_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; final FFN tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=bf16 |
| 110 | c36_bndall_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.69 GB | 0 | boundary_all=bf16 | Mixed manual probe: RYS entry, exit, and duplicate-boundary attention/SSM/FFN tensors at bf16.rules^blk\.(15|19|20|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta)|ffn_(down|gate|up))\.weight=bf16 |
| 111 | c10_union_rys_path_bf16_late_control_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.77 GB | 0 | ^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16; ^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0; ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 | Heavier union: RYS 15..24 pathway BF16 with late execution-control Q8_0.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 ^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 112 | c92_fa_q8_rysffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 18.82 GB | 0 | fa=q8_0, rys_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; RYS 15..24 FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q8_0 |
| 113 | c94_fa_q8_bndffn_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 19.02 GB | 0 | fa=q8_0, boundary_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; RYS entry, exit, and duplicate-boundary FFN tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=bf16 |
| 114 | c128_fa_q5k_outtok_q8_rysffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 19.06 GB | 0 | fa=q5_K, output_token=q8_0, rys_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K; token embedding and output tensors at q8_0; RYS 15..24 FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q8_0 |
| 115 | c134_fa_q6k_outtok_q8_rysffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 19.31 GB | 0 | fa=q6_K, output_token=q8_0, rys_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K; token embedding and output tensors at q8_0; RYS 15..24 FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q8_0 |
| 116 | c107_fa_bf16_lateffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 19.61 GB | 0 | fa=bf16, late_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; late FFN gate/down control tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 117 | c109_fa_bf16_finalffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 19.66 GB | 0 | fa=bf16, final_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; final FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=q8_0 |
| 118 | c103_fa_bf16_rysffn_q6k |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 19.95 GB | 0 | fa=bf16, rys_ffn=q6_K | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; RYS 15..24 FFN tensors at q6_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q6_K |
| 119 | c150_fa_bf16_bndffn_q8_lateffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 20.15 GB | 0 | fa=bf16, boundary_ffn=q8_0, late_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; RYS entry, exit, and duplicate-boundary FFN tensors at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 120 | c108_fa_bf16_lateffn_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 20.28 GB | 0 | fa=bf16, late_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; late FFN gate/down control tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=bf16 |
| 121 | c42_rysffn_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 20.40 GB | 0 | rys_ffn=bf16 | Mixed manual probe: RYS 15..24 FFN tensors at bf16.rules^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=bf16 |
| 122 | c110_fa_bf16_finalffn_bf16 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 20.41 GB | 0 | fa=bf16, final_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; final FFN tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=bf16 |
| 123 | c104_fa_bf16_rysffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 20.59 GB | 0 | fa=bf16, rys_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; RYS 15..24 FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q8_0 |
| 124 | c146_fa_bf16_outtok_q8_rysffn_q8 |
0.8146 | 0.8750 | 0.7328 | 0.8039 | 21.54 GB | 0 | fa=bf16, output_token=q8_0, rys_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; token embedding and output tensors at q8_0; RYS 15..24 FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q8_0 |
| 125 | c51_finalffn_q6k |
0.8117 | 0.8750 | 0.7261 | 0.8006 | 16.76 GB | 0 | final_ffn=q6_K | Mixed manual probe: final FFN tensors at q6_K.rules^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=q6_K |
| 126 | c86_fa_q6k_finalffn_bf16 |
0.8032 | 0.8750 | 0.7061 | 0.7906 | 18.18 GB | 0 | fa=q6_K, final_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K; final FFN tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^blk\.(60|64|68)\.ffn_(down|gate|up)\.weight=bf16 |
| 127 | c140_fa_q8_outtok_q8_rysffn_q8 |
0.8032 | 0.8750 | 0.7061 | 0.7906 | 19.77 GB | 0 | fa=q8_0, output_token=q8_0, rys_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; token embedding and output tensors at q8_0; RYS 15..24 FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q8_0 |
| 128 | c40_rysffn_q6k |
0.8031 | 0.8550 | 0.7328 | 0.7939 | 17.24 GB | 0 | rys_ffn=q6_K | Mixed manual probe: RYS 15..24 FFN tensors at q6_K.rules^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q6_K |
| 129 | c19_fav_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 16.59 GB | 0 | fa_v=q8_0 | Mixed manual probe: full-attention V tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_v\.weight=q8_0 |
| 130 | c43_lateattn_q6k |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 16.69 GB | 0 | late_attn=q6_K | Mixed manual probe: late full-attention control tensors at q6_K.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K |
| 131 | c29_dupsig_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 16.73 GB | 0 | dup_signal=q8_0 | Mixed manual probe: duplicated-side RYS 20..24 linear-attention and SSM signal path at q8_0.rules^blk\.(20|21|22|23|24)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 132 | c24_ryssig_q6k |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 16.73 GB | 0 | rys_signal=q6_K | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at q6_K.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q6_K |
| 133 | c11_fa_q5k |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 16.78 GB | 0 | fa=q5_K | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K |
| 134 | c44_lateattn_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 16.81 GB | 0 | late_attn=q8_0 | Mixed manual probe: late full-attention control tensors at q8_0.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 |
| 135 | c30_dupsig_bf16 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 17.05 GB | 0 | dup_signal=bf16 | Mixed manual probe: duplicated-side RYS 20..24 linear-attention and SSM signal path at bf16.rules^blk\.(20|21|22|23|24)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 |
| 136 | c38_bndffn_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 17.09 GB | 0 | boundary_ffn=q8_0 | Mixed manual probe: RYS entry, exit, and duplicate-boundary FFN tensors at q8_0.rules^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 |
| 137 | c71_fa_q5k_lateffn_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 17.14 GB | 0 | fa=q5_K, late_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K; late FFN gate/down control tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 138 | c08_late_control_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 17.17 GB | 0 | ^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0; ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 | Probe late execution-control retention: final five full-attention blocks plus late FFN gate/down tensors at Q8_0.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 139 | c59_tok_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 17.19 GB | 0 | token_only=q8_0 | Mixed manual probe: token embedding tensor at q8_0.rules^token_embd\.weight=q8_0 |
| 140 | c115_srcsig_bf16_dupsig_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 17.23 GB | 0 | src_signal=bf16, dup_signal=q8_0 | Mixed manual probe: source-side RYS 15..19 linear-attention and SSM signal path at bf16; duplicated-side RYS 20..24 linear-attention and SSM signal path at q8_0.rules^blk\.(15|16|17|18|19)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 ^blk\.(20|21|22|23|24)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 141 | c34_duppath_bf16 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 17.35 GB | 0 | dup_path=bf16 | Mixed manual probe: duplicated-side RYS 20..24 attention, linear-attention, and SSM path at bf16.rules^blk\.(20|21|22|23|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 |
| 142 | c132_fa_q5k_bndffn_q8_lateffn_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 17.67 GB | 0 | fa=q5_K, boundary_ffn=q8_0, late_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K; RYS entry, exit, and duplicate-boundary FFN tensors at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 143 | c66_fa_q5k_ryssig_bf16 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 17.78 GB | 0 | fa=q5_K, rys_signal=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K; RYS 15..24 linear-attention and SSM signal path at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 |
| 144 | c130_fa_q5k_outtok_q8_lateffn_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 18.08 GB | 0 | fa=q5_K, output_token=q8_0, late_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K; token embedding and output tensors at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 145 | c155_srcsig_bf16_dupsig_q8_outtok_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 18.17 GB | 0 | src_signal=bf16, dup_signal=q8_0, output_token=q8_0 | Mixed manual probe: source-side RYS 15..19 linear-attention and SSM signal path at bf16; duplicated-side RYS 20..24 linear-attention and SSM signal path at q8_0; token embedding and output tensors at q8_0.rules^blk\.(15|16|17|18|19)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 ^blk\.(20|21|22|23|24)\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 |
| 146 | c129_fa_q5k_outtok_q8_bndffn_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 18.26 GB | 0 | fa=q5_K, output_token=q8_0, boundary_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K; token embedding and output tensors at q8_0; RYS entry, exit, and duplicate-boundary FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 |
| 147 | c133_fa_q6k_outtok_q8_ryssig_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 18.32 GB | 0 | fa=q6_K, output_token=q8_0, rys_signal=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K; token embedding and output tensors at q8_0; RYS 15..24 linear-attention and SSM signal path at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 148 | c50_latectl_bf16 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 18.33 GB | 0 | late_control=bf16 | Mixed manual probe: late attention plus FFN gate/down control tensors at bf16.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=bf16 |
| 149 | c80_fa_q6k_rysffn_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 18.37 GB | 0 | fa=q6_K, rys_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q6_K; RYS 15..24 FFN tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q6_K ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q8_0 |
| 150 | c60_tok_bf16 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 18.38 GB | 0 | token_only=bf16 | Mixed manual probe: token embedding tensor at bf16.rules^token_embd\.weight=bf16 |
| 151 | c160_ryspath_q8_outtok_q8_lateffn_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 18.41 GB | 0 | rys_path=q8_0, output_token=q8_0, late_ffn=q8_0 | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at q8_0; token embedding and output tensors at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 152 | c157_srcpath_bf16_duppath_q8_outtok_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 18.58 GB | 0 | src_path=bf16, dup_path=q8_0, output_token=q8_0 | Mixed manual probe: source-side RYS 15..19 attention, linear-attention, and SSM path at bf16; duplicated-side RYS 20..24 attention, linear-attention, and SSM path at q8_0; token embedding and output tensors at q8_0.rules^blk\.(15|16|17|18|19)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 ^blk\.(20|21|22|23|24)\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 |
| 153 | c161_lateattn_bf16_lateffn_q8_outtok_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 18.61 GB | 0 | late_attn=bf16, late_ffn=q8_0, output_token=q8_0 | Mixed manual probe: late full-attention control tensors at bf16; late FFN gate/down control tensors at q8_0; token embedding and output tensors at q8_0.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 |
| 154 | c162_latectl_q8_outtok_q8_bndffn_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 18.65 GB | 0 | late_control=q8_0, output_token=q8_0, boundary_ffn=q8_0 | Mixed manual probe: late attention plus FFN gate/down control tensors at q8_0; token embedding and output tensors at q8_0; RYS entry, exit, and duplicate-boundary FFN tensors at q8_0.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=q8_0 |
| 155 | c139_fa_q8_outtok_q8_ryssig_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 18.78 GB | 0 | fa=q8_0, output_token=q8_0, rys_signal=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at q8_0; token embedding and output tensors at q8_0; RYS 15..24 linear-attention and SSM signal path at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q8_0 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 |
| 156 | c120_ryssig_bf16_rysffn_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 18.89 GB | 0 | rys_signal=bf16, rys_ffn=q8_0 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at bf16; RYS 15..24 FFN tensors at q8_0.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q8_0 |
| 157 | c124_latectl_bf16_outtok_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 19.27 GB | 0 | late_control=bf16, output_token=q8_0 | Mixed manual probe: late attention plus FFN gate/down control tensors at bf16; token embedding and output tensors at q8_0.rules^blk\.(52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=bf16 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 |
| 158 | c159_ryspath_bf16_outtok_q8_lateffn_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 19.45 GB | 0 | rys_path=bf16, output_token=q8_0, late_ffn=q8_0 | Mixed manual probe: RYS 15..24 attention, linear-attention, and SSM path at bf16; token embedding and output tensors at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(1[5-9]|2[0-4])\.(attn_(q|k|v|output)|attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 ^output\.weight=q8_0 ^token_embd\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 159 | c149_fa_bf16_ryssig_q8_lateffn_q8 |
0.8004 | 0.8750 | 0.6994 | 0.7872 | 19.96 GB | 0 | fa=bf16, rys_signal=q8_0, late_ffn=q8_0 | Mixed manual probe: all full-attention Q/K/V/O tensors at bf16; RYS 15..24 linear-attention and SSM signal path at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=bf16 ^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 160 | release_s010_iq4nl |
0.7954 | 0.8417 | 0.7328 | 0.7872 | 16.55 GB | 0 | ||
| 161 | c70_fa_q5k_bndffn_bf16 |
0.7952 | 0.8750 | 0.6872 | 0.7811 | 18.32 GB | 0 | fa=q5_K, boundary_ffn=bf16 | Mixed manual probe: all full-attention Q/K/V/O tensors at q5_K; RYS entry, exit, and duplicate-boundary FFN tensors at bf16.rules^blk\.(3|7|11|15|19|20|24|28|32|36|40|44|48|52|56|60|64|68)\.attn_(q|k|v|output)\.weight=q5_K ^blk\.(15|19|20|24)\.ffn_(down|gate|up)\.weight=bf16 |
| 162 | c158_ryssig_bf16_rysffn_q8_lateffn_q8 |
0.7906 | 0.8333 | 0.7328 | 0.7831 | 19.25 GB | 0 | rys_signal=bf16, rys_ffn=q8_0, late_ffn=q8_0 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at bf16; RYS 15..24 FFN tensors at q8_0; late FFN gate/down control tensors at q8_0.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=bf16 ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q8_0 ^blk\.(56|60|64|68)\.ffn_(down|gate)\.weight=q8_0 |
| 163 | c119_ryssig_q8_rysffn_q8 |
0.7889 | 0.8550 | 0.6994 | 0.7772 | 18.24 GB | 0 | rys_signal=q8_0, rys_ffn=q8_0 | Mixed manual probe: RYS 15..24 linear-attention and SSM signal path at q8_0; RYS 15..24 FFN tensors at q8_0.rules^blk\.(1[5-9]|2[0-4])\.(attn_qkv|attn_gate|ssm_(out|alpha|beta))\.weight=q8_0 ^blk\.(1[5-9]|2[0-4])\.ffn_(down|gate|up)\.weight=q8_0 |
Source accounting reproduced on page
| Dataset | Rows | What it represents |
|---|---|---|
| Stable promotion summary | 163 | Candidate rows from the stable 12-task hard-suite promotion screen. |
| Comprehensive summary | 440 | Later comprehensive candidate rows used for broader analysis and candidate-family comparison. |
| KritaLite | 11 | Real-world browser paint application harness rows. |
| SyncBoard | 8 | Real-world local-first board/state-management harness rows. |
| Legacy-codebase migration | 13 | Real-world existing-codebase memory migration harness rows. |
| Candidate catalog | 882 | Generated candidate recipe definitions used to annotate tested rows. Untested definitions are not counted as results. |