Publish gemma-4-26B-A4B-it distribution-fidelity artifact (part 6)

Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/attribution/routing/routing_00758.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b470d0bccc20ad72ad6676d07edec3f8aab5eb55714a1f59e3b7da95065170ab
+size 491608

Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/attribution/routing/routing_00759.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1110e4ab1780cae53ca97676822797d1526414f7474a314907cba20c339de20e
+size 491608

Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/attribution/routing/routing_00760.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:04a84be21c0c9ee1283f68ef3c71e862a8de28f922709f44a6aa5e1301a851ef
+size 491608

Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/attribution/routing/routing_00761.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e7954ec61394a85a70cb699a429b5804f61504861439310e7b86d1c7965e618e
+size 491608

Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/attribution/routing/routing_00762.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2511fd807c51f0e1606421c20d118e4c0e2f9b82739a6b766ff6b13fda60c558
+size 491608

Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/attribution/routing/routing_00763.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f8822ca33750d225bffccd017dca9ef065e17572744add7f8d10fda25a758660
+size 491608

Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/attribution/routing/routing_00764.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b779cd665458525be833699b0410d646c4d3ab86895ff45fb2e2d0e97ccf2298
+size 491608

Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/attribution/routing/routing_00765.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c153d4991edd679f6cd5e01f16aada000771f39de47cbaf8902b2161d94efab1
+size 491608

Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/attribution/routing/routing_00766.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a93ece76c557b61ddb06094f8f5add3e6aa0a5b65d7aa1ea7b7b674e9f716cd4
+size 491608

Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/attribution/routing/routing_00767.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:27010839450156046282d6c6fb18fb00fe5f23934ee0eaa8e9e8f1def971b4ff
+size 491608

Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/inspect.json

@@ -1,35 +1,35 @@
 {
+  "model": "/media/fmodels2/Neural-ICE/Gemma-4-26B-A4B-it-NVFP4",
+  "inspect_version": 4,
+  "weights_sha256": "e833541d24828cb56aa5190656ced93552a81d5e50f463b9e4a957d22ca15f32",
+  "weights_bytes": 16423205676,
+  "quant_method": "modelopt",
+  "declared": {},
+  "detected_scheme": "nvfp4",
+  "detected_block": 16,
+  "quant_algorithm": "modelopt",
   "coverage": {
-    "attention": {
-      "quantized": 115,
-      "weights": 115
-    },
-    "dense_mlp": {
-      "quantized": 90,
-      "weights": 90
+    "router": {
+      "weights": 30,
+      "quantized": 0
     },
     "experts": {
-      "quantized": 11520,
-      "weights": 11520
-    },
-    "router": {
-      "quantized": 0,
-      "weights": 30
+      "weights": 11520,
+      "quantized": 11520
     },
     "shared_expert": {
-      "quantized": 0,
-      "weights": 0
+      "weights": 0,
+      "quantized": 0
+    },
+    "attention": {
+      "weights": 115,
+      "quantized": 115
+    },
+    "dense_mlp": {
+      "weights": 90,
+      "quantized": 90
     }
   },
-  "declared": {},
-  "detected_block": 16,
-  "detected_scheme": "nvfp4",
-  "model": "/media/fmodels2/Neural-ICE/Gemma-4-26B-A4B-it-NVFP4",
-  "quant_algorithm": "modelopt",
-  "quant_method": "modelopt",
-  "quantized_names_sha256": "b1384f9775d673940254e21793af07954ca0b554e849cf4734ccb6712163dfa5",
   "unloadable_reason": null,
-  "weights_bytes": 16423205676,
-  "weights_bytes_source": "hub",
-  "weights_sha256": "16029a9fa15ce7791061bfc7b390034c904d1e6b91d0e0105f1300fd13bb80c4"
+  "quantized_names_sha256": "b1384f9775d673940254e21793af07954ca0b554e849cf4734ccb6712163dfa5"
 }

Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/report.md

@@ -1,12 +1,12 @@
 # gemma-4-26B-A4B-it / Neural-ICE-Gemma-4-26B-A4B-it-NVFP4: distribution fidelity
 
-**Mean KLD(reference || candidate) = 1.81298048** over 1572096 scored positions.
+**Mean KLD(reference || candidate) = 1.81880328** over 1572096 scored positions.
 
 | Mean | Median | p90 | p99 | Max | Top-1 agreement |
 |---|---|---|---|---|---|
-| 1.81298048 | 0.67060030 | 5.16405010 | 13.26047492 | 48.53921509 | 57.8623% |
+| 1.81880328 | 0.67534816 | 5.17001915 | 13.29482074 | 46.64394379 | 57.7526% |
 
-Reverse direction, KLD(candidate || reference): 2.10446417.
+Reverse direction, KLD(candidate || reference): 2.10499252.
 
 ## Identity
 
@@ -15,10 +15,10 @@ Reverse direction, KLD(candidate || reference): 2.10446417.
 | Reference checkpoint | /media/fmodels2/google/gemma-4-26B-A4B-it |
 | Reference config SHA-256 | ed0c1eb3633d |
 | Candidate checkpoint | /media/fmodels2/Neural-ICE/Gemma-4-26B-A4B-it-NVFP4 |
-| Candidate weights SHA-256 | 16029a9fa15ce779 |
+| Candidate weights SHA-256 | e833541d24828cb5 |
 | Suite | gemma4-31b-it-fidelity-1024x2048-v1 |
 | Suite token SHA-256 | 81d66409b44937fa |
-| Capture manifest SHA-256 | 395c569193357bb7 |
+| Capture manifest SHA-256 | d1f03dea74e9a9b9 |
 | Tokenizer | None |
 | Scored vocabulary | 262144 |
 | Declared vocabulary | 262144 |
@@ -30,16 +30,34 @@ Reverse direction, KLD(candidate || reference): 2.10446417.
 | Prefix caching | False |
 | max_num_seqs | 1 |
 | vLLM | 0.1.dev20446+gb2bc9171d |
-| vLLM commit | f4cee453bfd3 |
+| vLLM commit | c798225a7f19 |
+| vLLM dirty digest | e3b0c44298fc1c14 |
+| Compiled extensions | f2fbc7537b0f01f6 |
+| FlashInfer | 0.6.17 |
 | torch | 2.13.0+cu132 |
 | Driver | 580.173.02 |
 | GPUs | NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
-| Laws version | 9 |
+| Laws version | 12 |
 | Partition | analysis |
 
-## Component attribution
+## Routed-model intervention: QxQ and BxQ
 
-This reference routes tokens to experts, so the deployed mean is not a single effect. Each cell is the reference with one component rounded through the deployed scheme and run on BF16 kernels, scored on the same tokens against the same capture. The cells do not sum to the deployed mean and are not meant to: once a token is routed elsewhere, degrading the expert it no longer uses costs nothing.
+**The first axis is routing source; the second Q is the same unchanged quantized candidate in both runs.** QxQ uses the student's natural expert IDs. BxQ forces the BF16 teacher's expert IDs while the student computes its own gating weights for those experts.
+
+| Run | Routing | Mean KLD | Support |
+|---|---|---|---|
+| QxQ | student natural IDs; student gating weights | 1.81880328 | [report](report.json) |
+| BxQ | BF16 teacher IDs; student gating weights | 1.58869884 | [report](report.json) |
+| QxQ - BxQ | paired routing-intervention delta; not additive attribution | +0.23010445 |  |
+| Exact repeat | two QxQ, two forced-natural, two BxQ samples | certified |  |
+
+Natural QxQ selections changed from the teacher at **74.4082%** of (token, layer) choices. BxQ protocol 4; routing trace `65a995be0ed63037`.
+
+Natural control used exact-repeat certification: max position delta 0.000e+00 / 0.000e+00; absolute mean delta 0.000e+00 / 0.000e+00; natural route-repeat flips 0.000%.
+
+## Synthetic QDQ weight-rounding diagnostics
+
+These cells create synthetic BF16 checkpoints with selected weights quantize-dequantized through the deployed format, then route naturally on BF16 kernels. They diagnose weight rounding. They are not BxQ or QxQ, and they do not sum to either deployed run.
 
 No cell here holds expert selection fixed. Rounding any weight changes the residual stream, and every router downstream of that change sees different inputs, so each cell carries some routing movement of its own. What routing costs is measured directly under [Routing divergence](#routing-divergence), not inferred from these cells.
 
@@ -49,7 +67,7 @@ No cell here holds expert selection fixed. Rounding any weight changes the resid
 |---|---|
 | Scheme | nvfp4 |
 | Block | 16 |
-| Weights on disk | n/a |
+| Weights on disk | 15.30 GiB |
 | Declared method | modelopt |
 | Algorithm | modelopt |
 | Activation scheme | n/a |
@@ -61,27 +79,26 @@ No cell here holds expert selection fixed. Rounding any weight changes the resid
 | experts | 11520 / 11520 | all |
 | router | 0 / 30 | none |
 
-### Component cells
+### Synthetic QDQ cells
 
 | Cell | What it isolates | Mean KLD | Selections changed | Support |
 |---|---|---|---|---|
-| Experts only | expert weight precision | 0.65884572 | 50.5274% | [report](attribution/gemma-4-26B-A4B-it-qdq-experts-nvfp4.json) · [manifest](attribution/gemma-4-26B-A4B-it-qdq-experts-nvfp4-qdq-manifest.json) |
+| Experts only | expert weight precision | 0.66652247 | 50.5990% | [report](attribution/gemma-4-26B-A4B-it-qdq-experts-nvfp4.json) · [manifest](attribution/gemma-4-26B-A4B-it-qdq-experts-nvfp4-qdq-manifest.json) |
 | Router weights only | router weight precision, not the routing term | n/a — router ships unquantized | n/a | [routing divergence](#routing-divergence) |
-| Every quantized component | experts, attention, dense_mlp, BF16 kernels | 1.43550583 | 67.7501% | [report](attribution/gemma-4-26B-A4B-it-qdq-experts-attention-dense_mlp-nvfp4.json) · [manifest](attribution/gemma-4-26B-A4B-it-qdq-experts-attention-dense_mlp-nvfp4-qdq-manifest.json) |
-| Deployed | as shipped, quantized kernels | 1.81298048 | 74.4202% | [report.json](report.json) |
+| Every quantized component | experts, attention, dense_mlp, BF16 kernels | 1.43642041 | 67.7122% | [report](attribution/gemma-4-26B-A4B-it-qdq-experts-attention-dense_mlp-nvfp4.json) · [manifest](attribution/gemma-4-26B-A4B-it-qdq-experts-attention-dense_mlp-nvfp4-qdq-manifest.json) |
 
 Two terms sit outside every cell above, because no cell can hold them. They are the reason the deployed mean is not a precision measurement:
 
 | Term | Value | Share | What it is | Derived in |
 |---|---|---|---|---|
-| Beyond weight rounding | +0.37747464 | 21% of the deployed mean | activation quantization and kernel arithmetic | [Beyond weight rounding](#beyond-weight-rounding) |
-| Routing, measured | saturated | n/a | expert selection changed at 74.4202% of choices and every scored position rerouted, so no held-routing population remains to measure the excess against | [Routing divergence](#routing-divergence) |
+| Beyond weight rounding | +0.38238287 | 21% of the deployed mean | activation quantization and kernel arithmetic | [Beyond weight rounding](#beyond-weight-rounding) |
+| Routing, measured | saturated | n/a | expert selection changed at 74.4082% of choices and every scored position rerouted, so no held-routing population remains to measure the excess against | [Routing divergence](#routing-divergence) |
 
 The selections column is why none of these cells is routing-free: each one reroutes some tokens purely as a consequence of rounding weights upstream of a router it never touched.
 
 ### Beyond weight rounding
 
-**+0.37747464** (21% of the deployed mean). Every cell above rounds weights and runs on BF16 kernels. The deployed checkpoint differs from the composite cell by this much. The checkpoint quantizes weights only, so this term is kernel arithmetic. A weight-only analysis, which is what any QDQ cell is, cannot see it.
+**+0.38238287** (21% of the deployed mean). Every cell above rounds weights and runs on BF16 kernels. The deployed checkpoint differs from the composite cell by this much. The checkpoint quantizes weights only, so this term is kernel arithmetic. A weight-only analysis, which is what any QDQ cell is, cannot see it.
 
 The deployed checkpoint leaves the router in BF16, so router weight precision costs exactly nothing here. That is not the same as routing costing nothing: an identical router fed perturbed activations selects different experts, which is measured under [Routing divergence](#routing-divergence).
 
@@ -94,8 +111,8 @@ The same expert weights rounded through each scheme, which is the cell that disc
 | Scheme | Experts-only mean KLD | Variant | Support |
 |---|---|---|---|
 | fp8_block | not expressible |  |  |
-| mxfp8 | 0.46808904 | gemma-4-26B-A4B-it-qdq-experts-mxfp8 | [report](attribution/gemma-4-26B-A4B-it-qdq-experts-mxfp8.json) |
-| nvfp4 | 0.65884572 | gemma-4-26B-A4B-it-qdq-experts-nvfp4 | [report](attribution/gemma-4-26B-A4B-it-qdq-experts-nvfp4.json) |
+| mxfp8 | 0.46718718 | gemma-4-26B-A4B-it-qdq-experts-mxfp8 | [report](attribution/gemma-4-26B-A4B-it-qdq-experts-mxfp8.json) |
+| nvfp4 | 0.66652247 | gemma-4-26B-A4B-it-qdq-experts-nvfp4 | [report](attribution/gemma-4-26B-A4B-it-qdq-experts-nvfp4.json) |
 
 `fp8_block` has no rung on this model: fp8_block lays scales on a 128-wide grid and model.language_model.layers.0.experts.down_proj is 128x2816x704, whose 704 is not a multiple of 128. Padding the weight to fit the grid would change the scales, so the rung is reported absent rather than measured under a layout no checkpoint deploys.
 
@@ -103,19 +120,19 @@ Full cells, digests, and the deployed checkpoint's inspection are in [attributio
 
 ## Routing divergence
 
-Both runs saw the same tokens, and the candidate's routers are the reference's routers to the bit where the checkpoint leaves them unquantized. They still selected different experts, because the activations arriving at them were already perturbed. This section compares the two runs' recorded selections over 30 routed layers, 8 experts per token.
+This is the observed selection divergence in the normal QxQ run, not the BxQ intervention and not a QDQ router-weight cell. It compares the student's natural expert IDs with the teacher's recorded IDs over 30 routed layers, 8 experts per token.
 
 | Measure | Rate | Of what |
 |---|---|---|
-| Selections changed | 74.4202% | of 47162880 (token, layer) choices |
-| Expert slots changed | 41.1154% | of individual expert slots |
-| Highest-weighted expert changed | 26.7752% | includes reordering within an unchanged selection |
+| Selections changed | 74.4082% | of 47162880 (token, layer) choices |
+| Expert slots changed | 41.0994% | of individual expert slots |
+| Highest-weighted expert changed | 26.7558% | includes reordering within an unchanged selection |
 | Positions with any flip | 100.0000% | 1572096 of 1572096 scored positions |
 
 | Positions where | Mean KLD | Count |
 |---|---|---|
 | Routing held | n/a | 0 |
-| Routing flipped | 1.81298048 | 1572096 |
+| Routing flipped | 1.81880328 | 1572096 |
 
 **Ranking floor undefined.** Every one of the 1572096 scored positions rerouted in at least one layer, so there is no held-routing population to compare the flipped ones against. The floor is not small here, it is unmeasurable, and no rescore changes that: the perturbation is large enough that routing divergence is the whole picture rather than a correction to it. Rank against this candidate on the experts-only cell, and treat the deployed mean as unranked.
 
@@ -123,42 +140,42 @@ Divergence rises with how many layers rerouted, which is what distinguishes a si
 
 | Layers rerouted | Positions | Mean KLD |
 |---|---|---|
-| 4-7 | 16 | 2.22760546 |
-| 8+ | 1572080 | 1.81297626 |
+| 4-7 | 19 | 0.81431452 |
+| 8+ | 1572077 | 1.81881542 |
 
-Per-layer selection change runs from 28.34% to 90.57% (worst at layer 29), averaging 44.54% across the first quarter of routed layers and 88.40% across the last. Rerouting that grows with depth is accumulated perturbation, not router precision.
+Per-layer selection change runs from 28.33% to 90.55% (worst at layer 29), averaging 44.54% across the first quarter of routed layers and 88.37% across the last. Rerouting that grows with depth is accumulated perturbation, not router precision.
 
 ## Fidelity by domain
 
-The suite is stratified, so the mean above is an average over kinds of text that do not degrade equally. `x run` is a domain's mean divided by the run's, so 1.00 degrades exactly as much as the model overall. The reference's own top-1 probability is shown because a domain the reference finds harder will diverge more for that reason alone.
+The suite is stratified, so the mean above is an average over kinds of text that do not degrade equally. `deployed` is QxQ (natural student routing). `bxq` is the teacher-ID counterfactual on the same weights. `x run` is a domain's mean divided by the run's, so 1.00 degrades exactly as much as the model overall. The reference's own top-1 probability is shown because a domain the reference finds harder will diverge more for that reason alone.
 
 | Domain | Contexts | Reference top-1 | Mean KLD | x run | Worst context |
 |---|---|---|---|---|---|
-| Other multilingual content | 36 | 70.7% | 2.36055674 | 1.30 | 4.95748827 |
-| Natural dialogue, instruction following, and assistance | 96 | 77.3% | 2.08054197 | 1.15 | 3.92764610 |
-| Scientific and technical exposition | 96 | 69.8% | 1.99864352 | 1.10 | 3.68147474 |
-| Encyclopedic and factual reference | 96 | 68.9% | 1.96648130 | 1.08 | 3.25276868 |
-| Chinese across several content types | 72 | 63.3% | 1.87901146 | 1.04 | 3.03630672 |
-| News, history, economics, legal analysis, and essays | 72 | 68.9% | 1.81524354 | 1.00 | 4.17786204 |
-| Literary, narrative, and creative writing | 72 | 62.9% | 1.69993654 | 0.94 | 2.90391837 |
-| Worked mathematics, science, and formal reasoning | 96 | 72.4% | 1.60482712 | 0.89 | 2.87840228 |
-| Structured data, tool calls, APIs, JSON, and tables | 36 | 83.3% | 1.45054648 | 0.80 | 3.33826998 |
-| Source code, tests, technical documentation, and issue discussions | 96 | 84.2% | 1.37854253 | 0.76 | 2.93070775 |
+| Other multilingual content | 36 | 70.7% | 2.36378738 | 1.30 | 4.47943237 |
+| Natural dialogue, instruction following, and assistance | 96 | 77.4% | 2.07672787 | 1.14 | 4.15724368 |
+| Scientific and technical exposition | 96 | 69.9% | 2.07479035 | 1.14 | 3.38248339 |
+| Encyclopedic and factual reference | 96 | 68.7% | 2.02309414 | 1.11 | 3.37861824 |
+| Chinese across several content types | 72 | 63.0% | 1.84919395 | 1.02 | 2.94232623 |
+| News, history, economics, legal analysis, and essays | 72 | 68.9% | 1.82956264 | 1.01 | 3.43971685 |
+| Literary, narrative, and creative writing | 72 | 63.0% | 1.65811583 | 0.91 | 3.08165872 |
+| Worked mathematics, science, and formal reasoning | 96 | 72.3% | 1.56078164 | 0.86 | 3.25403348 |
+| Structured data, tool calls, APIs, JSON, and tables | 36 | 83.3% | 1.42163395 | 0.78 | 3.19270213 |
+| Source code, tests, technical documentation, and issue discussions | 96 | 84.2% | 1.39284493 | 0.77 | 2.87263039 |
 
 The same domains measured on the other attribution cells, which shows whether a domain's weakness is a property of the model or of the format:
 
-| Domain | deployed | mxfp8 | nvfp4 | all_quantized |
-|---|---|---|---|---|
-| Other multilingual content | 2.36055674 | 0.60547717 | 0.83417492 | 1.78787763 |
-| Natural dialogue, instruction following, and assistance | 2.08054197 | 0.59360268 | 0.81214064 | 1.63223190 |
-| Scientific and technical exposition | 1.99864352 | 0.49939479 | 0.69630321 | 1.64129071 |
-| Encyclopedic and factual reference | 1.96648130 | 0.48326683 | 0.70706639 | 1.60899741 |
-| Chinese across several content types | 1.87901146 | 0.46550453 | 0.68079090 | 1.42905706 |
-| News, history, economics, legal analysis, and essays | 1.81524354 | 0.46316973 | 0.63722594 | 1.41930026 |
-| Literary, narrative, and creative writing | 1.69993654 | 0.39760530 | 0.61395641 | 1.34757069 |
-| Worked mathematics, science, and formal reasoning | 1.60482712 | 0.39846297 | 0.55025835 | 1.22613314 |
-| Structured data, tool calls, APIs, JSON, and tables | 1.45054648 | 0.48035757 | 0.61317963 | 1.19883758 |
-| Source code, tests, technical documentation, and issue discussions | 1.37854253 | 0.36808736 | 0.51325925 | 1.10842929 |
+| Domain | deployed | bxq | mxfp8 | nvfp4 | all_quantized |
+|---|---|---|---|---|---|
+| Other multilingual content | 2.36378738 | 2.01060924 | 0.68445047 | 0.89229785 | 1.79519304 |
+| Natural dialogue, instruction following, and assistance | 2.07672787 | 1.79807095 | 0.59552252 | 0.82821448 | 1.66159067 |
+| Scientific and technical exposition | 2.07479035 | 1.82912511 | 0.50141256 | 0.70339667 | 1.68508002 |
+| Encyclopedic and factual reference | 2.02309414 | 1.73666327 | 0.49743044 | 0.71954225 | 1.56808539 |
+| Chinese across several content types | 1.84919395 | 1.67080916 | 0.44706156 | 0.68167667 | 1.40434265 |
+| News, history, economics, legal analysis, and essays | 1.82956264 | 1.57566891 | 0.44719073 | 0.64140586 | 1.41631587 |
+| Literary, narrative, and creative writing | 1.65811583 | 1.49461078 | 0.39173107 | 0.62338708 | 1.35504621 |
+| Worked mathematics, science, and formal reasoning | 1.56078164 | 1.36868653 | 0.37927146 | 0.52795210 | 1.23575505 |
+| Structured data, tool calls, APIs, JSON, and tables | 1.42163395 | 1.27786368 | 0.47775227 | 0.61546437 | 1.16853193 |
+| Source code, tests, technical documentation, and issue discussions | 1.39284493 | 1.18805084 | 0.36354693 | 0.52781126 | 1.09767677 |
 
 **Other multilingual content** is this candidate's weakest domain and **Source code, tests, technical documentation, and issue discussions** its strongest, a spread of 1.7x. A deployment weighted toward the weakest domain sees more divergence than the headline mean implies; the per-source breakdown and the reading are in [strata.md](strata.md) and [strata.json](strata.json).
 
@@ -166,9 +183,9 @@ The same domains measured on the other attribution cells, which shows whether a
 
 | Component | Value |
 |---|---|
-| Trunk (candidate hidden states, reference head) | 1.82570235 |
-| Deployed (candidate's own head) | 1.81298048 |
-| Head-associated delta (not additive) | -0.01272188 |
+| Trunk (candidate hidden states, reference head) | 1.81880333 |
+| Deployed (candidate's own head) | 1.81880328 |
+| Head-associated delta (not additive) | -0.00000005 |
 | Reference head | {'state': 'unquantized', 'workers': [{'heads': [{'name': 'language_model.lm_head', 'org_vocab_size': 262144, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}], 'logits_processor': {'head_dtype': 'torch.bfloat16', 'org_vocab_size': 262144, 'scale': 1.0, 'soft_cap': 30.0, 'type': 'LogitsProcessor', 'vocab_size': 262144}, 'primary': {'org_vocab_size': 262144, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}, 'state': 'unquantized'}]} |
 | Candidate head | {'runtime': {'state': 'quantized', 'workers': [{'heads': [{'name': 'language_model.lm_head', 'org_vocab_size': 262144, 'quant_method': 'UnquantizedLinearMethod', 'state': 'quantized', 'weight_dtype': 'torch.bfloat16'}], 'logits_processor': {'head_dtype': 'torch.bfloat16', 'org_vocab_size': 262144, 'scale': 1.0, 'soft_cap': 30.0, 'type': 'LogitsProcessor', 'vocab_size': 262144}, 'primary': {'org_vocab_size': 262144, 'quant_method': 'UnquantizedLinearMethod', 'state': 'quantized', 'weight_dtype': 'torch.bfloat16'}, 'state': 'quantized'}]}, 'state': 'quantized', 'static': {'ignored': True, 'lm_head_dtypes': {'model.language_model.embed_tokens.weight': 'BF16'}, 'lm_head_keys': [], 'output_weight_keys': ['model.language_model.embed_tokens.weight'], 'packed_keys': [], 'quant_method': 'modelopt', 'state': 'unquantized', 'tie_word_embeddings': True}} |
 
@@ -176,26 +193,26 @@ The same domains measured on the other attribution cells, which shows whether a
 
 | Position range | Positions | Mean KLD |
 |---|---|---|
-| 0–511 | 393216 | 1.47532770 |
-| 512–1023 | 393216 | 1.59163973 |
-| 1024–1535 | 393216 | 1.90300322 |
-| 1536–2046 | 392448 | 2.28286902 |
+| 0–511 | 393216 | 1.48458588 |
+| 512–1023 | 393216 | 1.59614646 |
+| 1024–1535 | 393216 | 1.90740159 |
+| 1536–2046 | 392448 | 2.28799558 |
 
 ## Error by reference confidence
 
 | Reference top-1 probability | Positions | Share | Mean KLD |
 |---|---|---|---|
-| [0.00, 0.25) | 113173 | 7.2% | 3.55195433 |
-| [0.25, 0.50) | 290937 | 18.5% | 2.67427173 |
-| [0.50, 0.75) | 308489 | 19.6% | 2.15030238 |
-| [0.75, 0.95) | 318287 | 20.2% | 1.71522176 |
-| [0.95, 1.00) | 541210 | 34.4% | 0.85155857 |
+| [0.00, 0.25) | 113465 | 7.2% | 3.55419825 |
+| [0.25, 0.50) | 291225 | 18.5% | 2.68340379 |
+| [0.50, 0.75) | 308315 | 19.6% | 2.15399698 |
+| [0.75, 0.95) | 317610 | 20.2% | 1.71816747 |
+| [0.95, 1.00) | 541481 | 34.4% | 0.85832234 |
 
 ## Top-K set agreement
 
 | K=1 | K=2 | K=3 | K=4 | K=5 |
 |---|---|---|---|---|
-| 57.8269% | 22.2254% | 6.8006% | 1.7973% | 0.4408% |
+| 57.7255% | 22.2352% | 6.7988% | 1.7936% | 0.4376% |
 
 ## Law compliance
 
@@ -207,17 +224,17 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | 2 | Determinism | PASS | eager enforced; prefix caching off, max_num_seqs=1 |
 | 3 | Frozen input | PASS | token hash matches suite gemma4-31b-it-fidelity-1024x2048-v1 [analysis]: 81d66409b44937fa |
 | 4 | Real vocabulary | PASS | scored 262144 real tokens of 262144 declared (0 padding rows) |
-| 5 | Manifest binding | PASS | all bound fields present; manifest 395c569193357bb7f32e750fa94d4156797bef41dc8f46f13370aca8764f4097 |
-| 6 | Provenance | PASS | torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ f4cee453bfd3, driver 580.173.02, 4 GPU(s) |
+| 5 | Manifest binding | PASS | all bound fields present; manifest d1f03dea74e9a9b9eaee1cba6a6131584c17e866198aa9660a55f0a79b3cf47a |
+| 6 | Provenance | PASS | torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ c798225a7f19, driver 580.173.02, 4 GPU(s) |
 | 7 | Storage integrity | PASS | hidden storage with bitwise-exact replay |
-| 8 | Head transparency | PASS | trunk 1.82570235, deployed 1.81298048, delta -0.012721877099676648 |
-| 9 | Tail and depth disclosure | PASS | mean 1.81298048, median 0.67060030, max 48.53921509, 4 depth buckets |
+| 8 | Head transparency | PASS | trunk 1.81880333, deployed 1.81880328, delta -5.164658944956102e-08 |
+| 9 | Tail and depth disclosure | PASS | mean 1.81880328, median 0.67534816, max 46.64394379, 4 depth buckets |
 | 10 | Comparability | PASS | comparability key fully resolved |
 | 11 | Freeze before qualification | NOT_APPLICABLE | partition is 'analysis' |
 | 12 | Reusable reference | PASS | suite, reference, head, and checksums present; published reference matches the scored capture on every bound field |
-| 14 | Component attribution | PASS | experts 0.65884572, router weights unquantized; selections changed at 74.420% of (token, layer) choices; ranking floor undefined: every scored position rerouted, so no held-routing population remains to compare against |
-| 15 | Domain disclosure | PASS | 10 domains disclosed; weakest other_multilingual at 2.36055674, strongest code_docs_issues at 1.37854253, spread 1.7x |
-| 16 | Candidate weight binding | PASS | scored weights 16029a9fa15ce779 as inspected |
+| 14 | Routed-model intervention | PASS | QxQ 1.81880328, BxQ 1.58869884, QxQ - BxQ +0.23010445; natural selections changed at 74.408% of (token, layer) choices; trace 65a995be0ed63037 |
+| 15 | Domain disclosure | PASS | 10 domains disclosed; weakest other_multilingual at 2.36378738, strongest code_docs_issues at 1.39284493, spread 1.7x |
+| 16 | Candidate weight binding | PASS | scored weights e833541d24828cb5 as inspected |
 | 13 | Recorded deviation | PASS | no overrides claimed |
 
 ## Environment
@@ -228,7 +245,7 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | Platform | Linux-6.8.0-137-generic-x86_64-with-glibc2.39 |
 | Python | 3.12.3 |
 | vLLM | 0.1.dev20446+gb2bc9171d |
-| vLLM commit | f4cee453bfd38b1afcb8e983517c385e426ec488 |
+| vLLM commit | c798225a7f197994ed502068fc6ac1ba8a9706f3 |
 | torch | 2.13.0+cu132 |
 | torch CUDA runtime | 13.2 |
 | cuDNN | 9.20.0 (92000) |
@@ -259,8 +276,18 @@ Credential values are never published. Set but redacted: `HF_TOKEN`.
 
 | Variable | Value |
 |---|---|
+| `CUBLAS_WORKSPACE_CONFIG` | `:4096:8` |
 | `CUDA_HOME` | `/usr/local/cuda-13.0` |
 | `HF_TOKEN` | `<redacted>` |
+| `NCCL_DETERMINISTIC` | `1` |
+| `PYTORCH_NVML_BASED_CUDA_CHECK` | `1` |
+| `TORCHINDUCTOR_CACHE_DIR` | `/tmp/torchinductor_phaedawg` |
+| `TORCHINDUCTOR_COMPILE_THREADS` | `1` |
+| `TRITON_CACHE_AUTOTUNING` | `1` |
+| `TRITON_PTXAS_BLACKWELL_PATH` | `/usr/local/cuda-13.0/bin/ptxas` |
+| `VLLM_BATCH_INVARIANT` | `1` |
+| `VLLM_MARLIN_USE_ATOMIC_ADD` | `0` |
+| `VLLM_MOE_USE_DEEP_GEMM` | `0` |
 
 ## Files in this artifact
 
@@ -268,26 +295,26 @@ Paths are relative to the artifact root, the same paths `checksums.txt` uses. Ve
 
 | Path | Size | What it is |
 |---|---|---|
-| `Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/report.md` | 20.28 KiB | This document. |
-| `Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/report.json` | 252.20 KiB | Every statistic behind it, machine-readable: per-bucket means, percentiles, agreement rates, and the phase timings. |
-| `Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/manifest.json` | 85.83 KiB | The capture manifest this result is bound to (Law 5). Scoring refuses to run if the live configuration differs from it. |
-| `Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/compliance.json` | 34.59 KiB | The law-by-law receipt, including the comparability key. |
-| `baselines/self-kld.json` | 4.76 KiB | The zero-baseline proof required by Law 1: the reference scored against a capture of itself. |
+| `Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/report.md` | 21.67 KiB | This document. |
+| `Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/report.json` | 564.40 KiB | Every statistic behind it, machine-readable: per-bucket means, percentiles, agreement rates, and the phase timings. |
+| `Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/manifest.json` | 87.32 KiB | The capture manifest this result is bound to (Law 5). Scoring refuses to run if the live configuration differs from it. |
+| `Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/compliance.json` | 370.41 KiB | The law-by-law receipt, including the comparability key. |
+| `baselines/self-kld.json` | 6.72 KiB | The zero-baseline proof required by Law 1: the reference scored against a capture of itself. |
 | `suite/suite-manifest.json` | 382.61 KiB | The frozen evaluation input's identity: token hashes per context and per partition, sources, strata, and the analysis/qualification split. |
 | `suite/tokens` | 17.32 MiB in 1024 files | The token IDs themselves. These are the evaluation input, not a description of it; retokenizing source text does not reproduce them. |
 | `suite/sources.json` | 737.48 KiB | Per-context provenance: dataset, revision, licence, source unit, and the deterministic token offset chosen within the document. |
 | `suite/validation/capability-overlap.json` | 1.48 KiB | The benchmark-contamination scan and every document it blocked. |
-| `reference/manifest.json` | 85.83 KiB | The reference capture's own manifest: geometry, vocabulary, storage mode, and a hash for every tensor file. |
+| `reference/manifest.json` | 87.32 KiB | The reference capture's own manifest: geometry, vocabulary, storage mode, and a hash for every tensor file. |
 | `reference` | 11.62 GiB in 771 files | The reusable reference distributions. Pass this directory as `--reference-logits` to score a new candidate against the same reference without loading the reference checkpoint. |
 | `reference/lm_head.safetensors` | 1.38 GiB | The reference language-model head, which turns the stored hidden states back into reference logits. |
-| `environment/runtime.json` | 2.26 KiB | Machine-readable provenance: torch, CUDA, cuDNN, NCCL, driver, devices, and the captured environment variables. |
+| `environment/runtime.json` | 5.33 KiB | Machine-readable provenance: torch, CUDA, cuDNN, NCCL, driver, devices, and the captured environment variables. |
 | `environment/summary.md` | 1.10 KiB | The same provenance as prose, plus an index of every captured file. |
 | `environment/toolchain-nvcc.txt` | 243 B | `nvcc --version` verbatim; `toolchain-gcc.txt` and `toolchain-ldd.txt` sit beside it. |
 | `environment/gpu-smi-query.txt` | 54.73 KiB | `nvidia-smi -q` verbatim: ECC state, persistence mode, clocks, and throttle reasons, any of which can move a bitwise result. |
-| `environment/pip-freeze.txt` | 4.31 KiB | Every installed package version in the scoring environment. |
-| `environment/models` | 11.35 KiB in 7 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
-| `checksums.txt` | 191.21 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
-| `LAWS.md` | 31.90 KiB | The laws this artifact was produced under, including the override procedure. |
+| `environment/pip-freeze.txt` | 4.47 KiB | Every installed package version in the scoring environment. |
+| `environment/models` | 25.42 KiB in 7 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
+| `checksums.txt` | 1.13 MiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
+| `LAWS.md` | 31.84 KiB | The laws this artifact was produced under, including the override procedure. |
 
 ## Scope
 

Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/strata.md

@@ -1,68 +1,68 @@
 # Fidelity by domain - gemma-4-26B-A4B-it / Neural-ICE-Gemma-4-26B-A4B-it-NVFP4
 
-768 contexts, 1572096 scored positions, mean 1.81298048, reference top-1 72.1%, top-1 agreement 57.8623%.
+768 contexts, 1572096 scored positions, mean 1.81880328, reference top-1 72.1%, top-1 agreement 57.7526%.
 
-`x run` is the domain's mean divided by the run's mean, so 1.00 is a domain that degrades exactly as much as the model as a whole. `Median ctx` and `p90 ctx` are the spread of per-context means within the domain.
+`x run` is the domain's mean divided by the run's mean, so 1.00 is a domain that degrades exactly as much as the model as a whole. The `deployed` cell is QxQ; `bxq` is the teacher-ID counterfactual on the same student weights. `Median ctx` and `p90 ctx` are the spread of per-context means within the domain.
 
 ## Ranked by domain
 
 | Domain | Contexts | Ref top-1 | Mean KLD | x run | Median ctx | p90 ctx | Worst ctx |
 |---|---|---|---|---|---|---|---|
-| Other multilingual content | 36 | 70.7% | 2.36055674 | 1.30 | 2.27779897 | 3.85267773 | 4.957488 |
-| Natural dialogue, instruction following, and assistance | 96 | 77.3% | 2.08054197 | 1.15 | 2.09957934 | 3.19753145 | 3.927646 |
-| Scientific and technical exposition | 96 | 69.8% | 1.99864352 | 1.10 | 1.97027905 | 2.83071522 | 3.681475 |
-| Encyclopedic and factual reference | 96 | 68.9% | 1.96648130 | 1.08 | 2.11264218 | 2.67482285 | 3.252769 |
-| Chinese across several content types | 72 | 63.3% | 1.87901146 | 1.04 | 1.87232546 | 2.74345675 | 3.036307 |
-| News, history, economics, legal analysis, and essays | 72 | 68.9% | 1.81524354 | 1.00 | 1.72994359 | 2.67870268 | 4.177862 |
-| Literary, narrative, and creative writing | 72 | 62.9% | 1.69993654 | 0.94 | 1.75721595 | 2.60320977 | 2.903918 |
-| Worked mathematics, science, and formal reasoning | 96 | 72.4% | 1.60482712 | 0.89 | 1.56024657 | 2.31793721 | 2.878402 |
-| Structured data, tool calls, APIs, JSON, and tables | 36 | 83.3% | 1.45054648 | 0.80 | 1.32311970 | 2.28654410 | 3.338270 |
-| Source code, tests, technical documentation, and issue discussions | 96 | 84.2% | 1.37854253 | 0.76 | 1.35580005 | 2.10252137 | 2.930708 |
+| Other multilingual content | 36 | 70.7% | 2.36378738 | 1.30 | 2.33727762 | 4.12772208 | 4.479432 |
+| Natural dialogue, instruction following, and assistance | 96 | 77.4% | 2.07672787 | 1.14 | 2.08255142 | 3.06389986 | 4.157244 |
+| Scientific and technical exposition | 96 | 69.9% | 2.07479035 | 1.14 | 2.14664769 | 2.87363265 | 3.382483 |
+| Encyclopedic and factual reference | 96 | 68.7% | 2.02309414 | 1.11 | 2.08843558 | 2.74405611 | 3.378618 |
+| Chinese across several content types | 72 | 63.0% | 1.84919395 | 1.02 | 1.87575851 | 2.60112117 | 2.942326 |
+| News, history, economics, legal analysis, and essays | 72 | 68.9% | 1.82956264 | 1.01 | 1.87272200 | 2.67503417 | 3.439717 |
+| Literary, narrative, and creative writing | 72 | 63.0% | 1.65811583 | 0.91 | 1.72611763 | 2.44466917 | 3.081659 |
+| Worked mathematics, science, and formal reasoning | 96 | 72.3% | 1.56078164 | 0.86 | 1.47363200 | 2.31385064 | 3.254033 |
+| Structured data, tool calls, APIs, JSON, and tables | 36 | 83.3% | 1.42163395 | 0.78 | 1.17168440 | 2.28769164 | 3.192702 |
+| Source code, tests, technical documentation, and issue discussions | 96 | 84.2% | 1.39284493 | 0.77 | 1.40792014 | 2.11998740 | 2.872630 |
 
 ## The same domains across the ladder
 
 Each column is a separate measurement of the same contexts, so a domain that moves between columns is one the format treats differently.
 
-| Domain | deployed | mxfp8 | nvfp4 | all_quantized |
-|---|---|---|---|---|
-| Other multilingual content | 2.36055674 | 0.60547717 | 0.83417492 | 1.78787763 |
-| Natural dialogue, instruction following, and assistance | 2.08054197 | 0.59360268 | 0.81214064 | 1.63223190 |
-| Scientific and technical exposition | 1.99864352 | 0.49939479 | 0.69630321 | 1.64129071 |
-| Encyclopedic and factual reference | 1.96648130 | 0.48326683 | 0.70706639 | 1.60899741 |
-| Chinese across several content types | 1.87901146 | 0.46550453 | 0.68079090 | 1.42905706 |
-| News, history, economics, legal analysis, and essays | 1.81524354 | 0.46316973 | 0.63722594 | 1.41930026 |
-| Literary, narrative, and creative writing | 1.69993654 | 0.39760530 | 0.61395641 | 1.34757069 |
-| Worked mathematics, science, and formal reasoning | 1.60482712 | 0.39846297 | 0.55025835 | 1.22613314 |
-| Structured data, tool calls, APIs, JSON, and tables | 1.45054648 | 0.48035757 | 0.61317963 | 1.19883758 |
-| Source code, tests, technical documentation, and issue discussions | 1.37854253 | 0.36808736 | 0.51325925 | 1.10842929 |
+| Domain | deployed | bxq | mxfp8 | nvfp4 | all_quantized |
+|---|---|---|---|---|---|
+| Other multilingual content | 2.36378738 | 2.01060924 | 0.68445047 | 0.89229785 | 1.79519304 |
+| Natural dialogue, instruction following, and assistance | 2.07672787 | 1.79807095 | 0.59552252 | 0.82821448 | 1.66159067 |
+| Scientific and technical exposition | 2.07479035 | 1.82912511 | 0.50141256 | 0.70339667 | 1.68508002 |
+| Encyclopedic and factual reference | 2.02309414 | 1.73666327 | 0.49743044 | 0.71954225 | 1.56808539 |
+| Chinese across several content types | 1.84919395 | 1.67080916 | 0.44706156 | 0.68167667 | 1.40434265 |
+| News, history, economics, legal analysis, and essays | 1.82956264 | 1.57566891 | 0.44719073 | 0.64140586 | 1.41631587 |
+| Literary, narrative, and creative writing | 1.65811583 | 1.49461078 | 0.39173107 | 0.62338708 | 1.35504621 |
+| Worked mathematics, science, and formal reasoning | 1.56078164 | 1.36868653 | 0.37927146 | 0.52795210 | 1.23575505 |
+| Structured data, tool calls, APIs, JSON, and tables | 1.42163395 | 1.27786368 | 0.47775227 | 0.61546437 | 1.16853193 |
+| Source code, tests, technical documentation, and issue discussions | 1.39284493 | 1.18805084 | 0.36354693 | 0.52781126 | 1.09767677 |
 
 ## Ranked by source dataset
 
 | Source | Contexts | Ref top-1 | Mean KLD | x run | Median ctx | p90 ctx | Worst ctx |
 |---|---|---|---|---|---|---|---|
-| wikipedia_ru | 6 | 68.3% | 3.25179235 | 1.79 | 3.32979054 | 4.20144500 | 4.201445 |
-| wikipedia_cs | 6 | 70.6% | 3.13637570 | 1.73 | 3.38542783 | 4.95748827 | 4.957488 |
-| wikipedia_de | 7 | 77.2% | 2.10556411 | 1.16 | 2.25848982 | 2.36777393 | 2.367774 |
-| wikipedia_fr | 3 | 69.4% | 2.10499370 | 1.16 | 2.17486815 | 2.41453792 | 2.414538 |
-| wildchat | 96 | 77.3% | 2.08054197 | 1.15 | 2.09957934 | 3.19753145 | 3.927646 |
-| open_news | 23 | 66.5% | 2.03162879 | 1.12 | 1.87281991 | 3.27167408 | 4.177862 |
-| scientific_papers | 96 | 69.8% | 1.99864352 | 1.10 | 1.97027905 | 2.83071522 | 3.681475 |
-| wikipedia_zh | 39 | 62.4% | 1.99365703 | 1.10 | 2.16009428 | 2.75896891 | 2.846686 |
-| wikipedia_en | 96 | 68.9% | 1.96648130 | 1.08 | 2.11264218 | 2.67482285 | 3.252769 |
-| wikipedia_ja | 7 | 69.8% | 1.90502534 | 1.05 | 1.99524572 | 2.29983517 | 2.299835 |
-| public_domain_review | 26 | 64.7% | 1.83727841 | 1.01 | 1.91656603 | 2.70666299 | 2.938417 |
-| wikipedia_es | 7 | 67.9% | 1.75170385 | 0.97 | 1.65127795 | 2.83761663 | 2.837617 |
-| wikisource_zh | 33 | 64.5% | 1.74352124 | 0.96 | 1.71493221 | 2.42909737 | 3.036307 |
-| public_domain_books | 72 | 62.9% | 1.69993654 | 0.94 | 1.75721595 | 2.60320977 | 2.903918 |
-| libretexts | 96 | 72.4% | 1.60482712 | 0.89 | 1.56024657 | 2.31793721 | 2.878402 |
-| regulations | 23 | 76.1% | 1.57394931 | 0.87 | 1.65594901 | 2.35188648 | 3.129034 |
-| starcoder_structured | 36 | 83.3% | 1.45054648 | 0.80 | 1.32311970 | 2.28654410 | 3.338270 |
-| stackv2 | 44 | 80.0% | 1.40411448 | 0.77 | 1.44263735 | 2.09550064 | 2.520085 |
-| github_code | 52 | 87.8% | 1.35690474 | 0.75 | 1.22107401 | 2.14591189 | 2.930708 |
+| wikipedia_ru | 6 | 67.4% | 3.42883253 | 1.89 | 3.58539667 | 4.24686247 | 4.246862 |
+| wikipedia_cs | 6 | 70.9% | 2.94733312 | 1.62 | 2.91655752 | 4.47943237 | 4.479432 |
+| wikipedia_fr | 3 | 69.8% | 2.45964151 | 1.35 | 2.38770855 | 3.08783978 | 3.087840 |
+| wikipedia_de | 7 | 77.8% | 2.14317143 | 1.18 | 2.13962176 | 2.53813189 | 2.538132 |
+| wildchat | 96 | 77.4% | 2.07672787 | 1.14 | 2.08255142 | 3.06389986 | 4.157244 |
+| scientific_papers | 96 | 69.9% | 2.07479035 | 1.14 | 2.14664769 | 2.87363265 | 3.382483 |
+| open_news | 23 | 66.7% | 2.06663665 | 1.14 | 1.99121440 | 3.14055687 | 3.383914 |
+| wikipedia_en | 96 | 68.7% | 2.02309414 | 1.11 | 2.08843558 | 2.74405611 | 3.378618 |
+| wikipedia_zh | 39 | 61.8% | 1.98919047 | 1.09 | 2.09188728 | 2.72111365 | 2.939617 |
+| wikipedia_ja | 7 | 69.6% | 1.89628395 | 1.04 | 1.82755000 | 2.69385492 | 2.693855 |
+| public_domain_review | 26 | 64.5% | 1.79542688 | 0.99 | 1.81613992 | 2.67503417 | 3.235695 |
+| wikisource_zh | 33 | 64.5% | 1.68374352 | 0.93 | 1.68374639 | 2.24469261 | 2.942326 |
+| public_domain_books | 72 | 63.0% | 1.65811583 | 0.91 | 1.72611763 | 2.44466917 | 3.081659 |
+| regulations | 23 | 75.9% | 1.63107689 | 0.90 | 1.70736205 | 2.39118263 | 3.439717 |
+| wikipedia_es | 7 | 67.8% | 1.59774849 | 0.88 | 1.31275107 | 2.78642330 | 2.786423 |
+| libretexts | 96 | 72.3% | 1.56078164 | 0.86 | 1.47363200 | 2.31385064 | 3.254033 |
+| stackv2 | 44 | 79.9% | 1.44362618 | 0.79 | 1.54038322 | 2.18371791 | 2.872630 |
+| starcoder_structured | 36 | 83.3% | 1.42163395 | 0.78 | 1.17168440 | 2.28769164 | 3.192702 |
+| github_code | 52 | 87.9% | 1.34987618 | 0.74 | 1.26600821 | 2.07561569 | 2.544324 |
 
 ## Reading
 
-- Weakest domain: **Other multilingual content** at 2.36055674, 1.30x the run mean of 1.81298048 over 36 context(s).
-- Strongest domain: **Source code, tests, technical documentation, and issue discussions** at 1.37854253, 0.76x the run mean. The spread across domains is 1.7x.
+- Weakest domain: **Other multilingual content** at 2.36378738, 1.30x the run mean of 1.81880328 over 36 context(s).
+- Strongest domain: **Source code, tests, technical documentation, and issue discussions** at 1.39284493, 0.77x the run mean. The spread across domains is 1.7x.
 - Every cell is weakest on the same domain (Other multilingual content), so the ranking is a property of the model and the text rather than of one format.
 

QXQ.md

@@ -0,0 +1,356 @@
+# QxQ and BxQ — Routed-Model Distribution Fidelity
+
+**Applies to:** laws version 12, Law 14.
+**Read first:** [`LAWS.md`](LAWS.md) for the laws, [`README.md`](README.md) for the
+harness.
+
+This document explains what the QxQ and BxQ cells measure, why measuring them at
+all requires bitwise-exact repeat from every kernel on the path, which MoE
+backends have earned the right to produce a published number, and how the
+pipeline refuses the ones that have not.
+
+## 1. Why one number is not enough for a routed model
+
+For a dense model, quantization error is weight rounding. One KLD against the
+BF16 reference describes it, and the only question left is how it distributes
+across text.
+
+A Mixture-of-Experts model breaks that. Two unrelated things happen at once, and
+a single mean hides both:
+
+1. The experts round their weights, exactly as a dense layer would.
+2. The router sends tokens to *different experts* than the reference would,
+   because the activations arriving at the router are themselves quantized.
+
+These have opposite consequences. Rounding is a fidelity tax that shrinks
+predictably with bit width. Routing divergence means the model is evaluating a
+different function — a different subnetwork per token — and it does not shrink
+predictably with anything. A candidate whose loss is mostly rounding and a
+candidate whose loss is mostly rerouting can post identical means and behave
+differently in deployment. Ranking them on that shared mean is the error Law 14
+exists to prevent.
+
+## 2. The two cells
+
+Both cells score the same candidate against the same BF16 teacher on the same
+frozen token suite, and differ only in who chooses the experts.
+
+**QxQ** is the candidate exactly as deployed: quantized weights, and its own
+router choosing its own experts. This is the number that describes what a user
+would actually run, and it is what routed candidates are ranked on.
+
+**BxQ** replays the same tokens with the teacher's expert IDs forced into the
+student, while the student keeps its own gating *weights*. Forcing the IDs
+removes the routing disagreement and leaves the rounding, so BxQ is the fidelity
+the candidate would have if its router agreed with the reference.
+
+The keeping of student gating weights is deliberate and easy to get wrong. If
+BxQ also took the teacher's gate values it would stop being a measurement of the
+student's experts and become a partial reconstruction of the teacher. Only the
+selection is intervened on; the weighting stays the candidate's own.
+
+**QxQ − BxQ** is therefore the routing contribution, reported as the paired
+intervention delta. The **natural route flip rate** reports how often the two
+disagree at all, measured over `(token, layer)` choices on the natural QxQ run —
+not on the intervened one, where by construction there is nothing to count.
+
+A worked example from the gemma-4-26B-A4B-it family: `Intel/gemma-4-26B-A4B-it-int4-AutoRound`
+scores QxQ 1.17816288 and BxQ 0.90308200, a delta of +0.27508088, and flips
+93.07% of natural routing choices. It sits mid-pack on deployed fidelity, but the
+decomposition says something no mean could: nearly all of its loss is the router,
+not the rounding. Its neighbours in the same group carry deltas near +0.15 at
+comparable flip rates.
+
+## 3. Why the subtraction demands bitwise-exact repeat
+
+QxQ − BxQ is a difference between two separate forward passes. If scoring the
+same tokens twice does not reproduce the same logits bit for bit, then a delta of
++0.15 is indistinguishable from kernel noise, and the decomposition is
+storytelling.
+
+So the pipeline does not assume reproducibility, it measures it. Every routed
+candidate is scored, replayed, and the two results compared; the published
+`repeat_delta` is that difference and it is required to be exactly `0.0`. A
+candidate whose backend cannot deliver that is marked uncertified for exact
+repeat and cannot be published as a Law 14 measurement.
+
+This is the entire reason the effort behind this document was determinism work
+rather than metric work. The metric is a subtraction; the difficulty is earning
+the right to subtract.
+
+## 4. How MoE kernels break exact repeat
+
+Batch invariance is the property that a token's output does not depend on which
+other tokens share its batch, nor on their order. Most MoE kernels violate it,
+and not by accident: grouping tokens by expert and reducing partial sums in
+whatever order the grouping produced is the fast way to do it. That is a correct
+optimization for serving, where nobody compares two runs, and it is fatal here.
+
+Four fixes were required before routed models could be scored at all.
+
+**Marlin token ordering and reduction.** The Marlin MoE path was ported to a
+canonical token order with a full-K reduction, which is what makes
+`MarlinExperts` and `BatchedMarlinExperts` batch invariant. This carries the bulk
+of the field: every int4 AWQ, GPTQ, and W4A16 routed candidate scores here.
+
+**Qwen GDN attention.** Certified only on its NVIDIA CUDA, non-speculative,
+per-sequence path, with FlashInfer GDN context parallelism disabled. Without this
+the Qwen3.6 family could not run under `VLLM_BATCH_INVARIANT` at all.
+
+**Dense NVFP4 linear layers.** Weight-only W4A16 NVFP4 linear layers use
+deterministic emulation, because dense Marlin is not batch invariant.
+
+**Expert parallelism is never certified.** An EP path reduces across ranks in
+completion order. No allowlist entry overrides this.
+
+## 5. Certification fails closed
+
+A kernel's `_supports_batch_invariance()` is a claim about itself. It is not
+evidence, and at least one kernel's claim is false in a way that silently
+destroys the measurement (§6).
+
+`inspect_model_moe_backends` in `vllm/v1/sample/kld.py` therefore grants
+`certified_for_exact_repeat` only when all three of these hold:
+
+1. The expert class is named in `_EXACT_REPEAT_CERTIFIED_EXPERTS`, an allowlist
+   earned by passing an exact-repeat probe on real suite content. The report
+   carries this separately as `exact_repeat_probed`.
+2. The kernel also self-declares batch invariance.
+3. The run is not expert-parallel.
+
+The allowlist currently holds `MarlinExperts`, `BatchedMarlinExperts`,
+`TritonExperts`, and `Nvfp4QuantizationEmulationTritonExperts`.
+
+The default is refusal. An unprobed backend — including one that arrives with a
+future vLLM bump — is uncertified until somebody probes it, and an uncertified
+backend is reported as uncertified rather than published as a number. This
+matters more than it sounds: the failure mode it replaces is a kernel quietly
+producing garbage that the harness dutifully writes down as a fidelity result.
+
+DeepGEMM, FlashInfer MoE, AITER, XPU, CPU, and every expert-parallel path remain
+uncertified.
+
+## 6. Case study: a kernel that declares batch invariance and returns NaN
+
+`CutlassExpertsFp4` self-declares batch invariance. Under
+`VLLM_BATCH_INVARIANT=1`, on W4A4 NVFP4 checkpoints, it takes finite inputs to
+NaN. The scorer then refuses a non-finite KLD and the engine dies, which is how
+this surfaced: two NVFP4 candidates failing with `KLD is not finite`, no report
+produced, and nothing to indicate the kernel rather than the checkpoint.
+
+Localizing it took two purpose-built tools.
+
+`scripts/scan_checkpoint_nonfinite.py` reads a checkpoint's tensors and counts
+non-finite values, zero scales, and per-row magnitude outliers in scale tensors.
+`torch.isfinite` has no CPU kernel for several float8 dtypes, so `nonfinite_mask`
+tests FP8 bit patterns directly — `(bits & 0x7F) == 0x7F` for `e4m3fn`, and
+`>= 0x7C` for `e5m2`. The scan came back clean, which established that the NaN
+was generated at runtime rather than baked into the weights, and moved the
+investigation off the checkpoint.
+
+`scripts/nan_first_module_probe.py` installs forward hooks and reports the first
+module whose output goes non-finite from finite inputs, together with the first
+row index at which it happens. Two capabilities were decisive. `--context-file`
+replays a real suite context instead of a synthetic prompt, because synthetic
+prompts are repetitive and never route diversely enough to trigger the bug.
+`--prompt-logprobs` computes logits at every prompt position, which is what the
+scoring harness does and what a generation-only probe does not.
+
+With both, the failure is exactly reproducible: `moe.experts` at layer 0, row 814
+of `context-0002`, and layer 14, row 170 of `context-0004`.
+
+The `--moe-backend` override then isolated the kernel:
+
+| `VLLM_BATCH_INVARIANT` | Backend | Result |
+| --- | --- | --- |
+| 0 | `FLASHINFER_CUTLASS` (auto) | pass |
+| 1 | `VLLM_CUTLASS` (auto) | **NaN** |
+| 1 | `marlin` (forced) | pass |
+| 1 | `emulation` (forced) | pass |
+
+The bug is specific to the CUTLASS FP4 experts under batch invariance, not to
+NVFP4, not to the checkpoints, and not to batch invariance itself. It is why §5
+fails closed: this kernel's self-declaration was the only thing standing between
+a broken forward pass and a published fidelity number.
+
+## 7. W4A4 NVFP4 scores on faithful emulation
+
+A W4A16 export and a W4A4 export of the same model differ only in whether
+activations are also quantized, and that difference decides which kernels can
+score the checkpoint honestly. A repository name says nothing reliable about it,
+so `_quantizes_activations_to_fp4` in `examples/offline_inference/score_mode_kld.py`
+reads the checkpoint: any entry in `quantization_config.config_groups` whose
+`input_activations` declares `num_bits: 4` and `type: "float"` makes it W4A4.
+
+For those checkpoints the scorer pins `moe_backend="emulation"`, which
+quantize-dequantizes both activation stages in the checkpoint's own scheme and is
+batch invariant by construction.
+
+Two details matter and both were bugs first:
+
+**The pin belongs on the student only.** It is applied to `student_kwargs`, not
+the shared `llm_kwargs`. Applied to the latter it propagates to the unquantized
+BF16 teacher, which has no such scheme, and reference engine initialization fails.
+
+**Marlin is refused here, not merely not preferred.** Marlin is batch invariant
+and would run, but its MoE path drops activation scales — it would score a W4A4
+checkpoint as though it were W4A16 and report a fidelity the checkpoint never
+delivers. A wrong number that passes every law is worse than a refusal.
+
+The honest consequence, stated in the published card: an NVFP4 result measures
+the quantization scheme rather than a native FP4 kernel's own rounding. Read the
+family's leaderboard with that in mind. Two NVFP4 exports in the
+gemma-4-26B-A4B-it family carry byte-identical QDQ diagnostics (0.66652247 and
+1.43642041) and yet score QxQ 1.16299506 and 1.75580190 — a 0.6 nat spread that
+lives entirely in the activation scheme and kernel path, not in the weights.
+
+## 8. Checkpoint defects the pipeline had to fix, not tolerate
+
+Determinism gets you a repeatable number. It does not get you a *correct* one if
+the checkpoint is being loaded wrong, and two classes of loading bug were found
+by scoring rather than by tests.
+
+**AutoRound int4 geometry.** AutoRound exports use group sizes that do not divide
+the input dimension, producing a partial final group. The packed-row and
+scale-group arithmetic — `scales_size`, `num_groups`, `size_k`, and the encoded
+symmetric zero used to pad — was wrong at exactly those boundaries, in both the
+Marlin and AutoGPTQ paths. A read-only `safetensors` header probe established the
+real geometry before any code changed; assumptions about row counts had already
+cost several wrong fixes. Regression tests cover the padded reduction allocation,
+the encoded-zero padding, single-rank partial groups, and the dense path's packed
+rows with a partial scale group.
+
+**A quantized router with nowhere to land.** The AutoRound Gemma-4 export
+quantizes the router projection. vLLM's `Gemma4Router` hardcodes its `GateLinear`
+as unquantized BF16, so the checkpoint's `qweight`, `qzeros`, and `scales` had no
+destination, no loader complained, and `router.proj.weight` was left as
+uninitialized memory. The router emitted NaN from finite inputs and KLD went
+non-finite.
+
+Note how this presents: identical symptom to §6, entirely different cause. The
+module probe is what separated them, naming `layers.0.router.proj` rather than
+`moe.experts`. The fix dequantizes those tensors into BF16 at load time in
+`Gemma4Model.load_weights`, with tests in `tests/kernels/moe/test_gemma4router.py`
+covering the transposed packed values and the refusal of an indivisible group
+count. That checkpoint now scores 1.17816288 and passes all sixteen laws.
+
+## 9. The runtime the numbers are bound to
+
+Scoring pins `VLLM_BATCH_INVARIANT=1`, disables DeepGEMM
+(`VLLM_MOE_USE_DEEP_GEMM=0`) and FlashInfer autotune, sets `NCCL_DETERMINISTIC=1`
+and `CUBLAS_WORKSPACE_CONFIG=:4096:8`, enforces eager execution, disables prefix
+caching, and holds `max_num_seqs=1`.
+
+None of that is optional and none of it is a performance setting. Each one closes
+a path by which two runs of the same tokens could diverge.
+
+Because that runtime *is* part of the result, its identity is bound into the
+comparability key: `vllm_commit`, `vllm_dirty_digest`,
+`compiled_extensions_sha256`, `torch`, `driver`, and `gpu_names`, alongside the
+suite and geometry. Two candidates are ranked against each other only when all of
+it matches. A consequence worth internalizing before committing to this
+repository: **any commit changes `vllm_commit`, and a dirty tree changes
+`vllm_dirty_digest`, so the next scoring run treats every prior report as stale
+and rescores the family.** That is correct behaviour, not a bug, and it is why
+harness fixes are best batched.
+
+## 10. Gates that stop a correct measurement from being published wrong
+
+Each of the following was found by a law refusing to publish, at the end of a
+multi-hour campaign, rather than by a test. They are recorded here because the
+failure mode is characteristic: the measurement is fine, and the metadata binding
+it to a suite, a geometry, or a reference capture is not.
+
+**Suite-driven geometry.** `_expected_geometry` derives the expected row count and
+context length from the suite manifest and the active partition, not from the
+dataset-driven `rows` and `context_length` on the config. Those config fields
+describe a dataset-driven run only; reading them for a suite-driven one compared
+768 real rows against a config default of 1024 and failed every candidate.
+
+**A field nobody wrote.** `context_length` was absent from every report, so a gate
+comparing it always failed. The writer in `score_mode_kld.py` now records it.
+Existing reports were backfilled from their bound capture manifests after
+verifying the recorded `capture_manifest_sha256`, refusing any mismatch.
+
+**Captures that moved underneath a report.** Assembly publishes the reference
+capture beside the report that cites it. If the capture is rebuilt after a report
+scores, the pair cites a manifest nothing hashes to, and Laws 5 and 14 refuse it —
+after the whole campaign. `_score_report_is_current` now treats a changed capture
+manifest as staleness, where a rescore is cheap.
+
+**Currency must not depend on processing order.** The gate above compares against
+whatever capture is on disk at that instant, and silently passes when the file is
+absent. That is order-dependent: early candidates in a run match the capture their
+predecessor left behind and are skipped, then a later rescore replaces it and
+strands them. `_candidate_complete` therefore delegates to the same
+`_score_report_is_current` the scorer uses, which compares each report's recorded
+commit against the live runtime and cannot be defeated by ordering.
+
+The last one has a worked example. In the gemma-4-26B-A4B-it family, one candidate
+was skipped early in a run on a stale `ec11b8…` capture, while a later rescore
+rebuilt the capture as `d1f03dea…` on a new commit. Nine candidates published on
+the new commit; the tenth kept a report from `fdc0b57e` bound to a capture the
+family no longer published. It failed Law 12 alone and ranked in a comparability
+group of one — every other law passed, because they read the candidate's own
+manifest. It was withdrawn via `excluded_candidates`, which records the repo,
+revision, and reason in `excluded-candidates.json` beside the family, rather than
+being dropped from the config silently.
+
+## 11. Reproducing and extending
+
+Certify a new MoE backend before trusting it. In order:
+
+```bash
+# 1. Is the NaN in the weights, or generated at runtime?
+python scripts/scan_checkpoint_nonfinite.py --model <checkpoint> --stats
+
+# 2. Which module goes non-finite first, on real content, at which row?
+VLLM_BATCH_INVARIANT=1 python scripts/nan_first_module_probe.py \
+    --model <checkpoint> --context-file <suite>/contexts/context-0002.json \
+    --prompt-logprobs
+
+# 3. Is it the kernel? Re-run pinning each backend in turn.
+VLLM_BATCH_INVARIANT=1 python scripts/nan_first_module_probe.py \
+    --model <checkpoint> --context-file <suite>/contexts/context-0002.json \
+    --prompt-logprobs --moe-backend marlin
+
+# 4. One candidate end to end: finite KLD, and repeat exactly 0.
+python fidelity/campaign.py smoke --config <campaign>.json \
+    --only-candidate <name>
+
+# 5. The family.
+python fidelity/campaign.py all --config <campaign>.json
+```
+
+A backend that passes the probe on real content, at full context, with prompt
+logprobs on, may be added to `_EXACT_REPEAT_CERTIFIED_EXPERTS`. Nothing else
+qualifies it, and self-declaration never does.
+
+Note that `--only-candidate` applies to `smoke` only. To rescore a single
+candidate of an assembled family, delete its report from
+`<work>/reports/<tag>.json`; the completeness gate then rescores exactly that one
+and skips the rest.
+
+## 12. What these numbers do not say
+
+**They are not comparable outside their group.** Not against numbers from another
+suite, geometry, runtime, or laws version, and not against any published
+elsewhere. The comparability key is printed with every leaderboard group for
+exactly this reason.
+
+**An NVFP4 result measures the scheme, not a native FP4 kernel.** See §7. The
+faithful-emulation path is a deliberate substitution, disclosed rather than hidden.
+
+**The QDQ ladder is diagnostic, never a candidate.** Those cells round weights on
+synthetic BF16 checkpoints and route naturally. They are not QxQ or BxQ, they are
+not rankable against deployed candidates, and the published tables separate them.
+
+**A flip rate is not an error rate.** Two experts disagreeing on a token is not
+by itself a wrong answer; the delta is what quantifies the cost. AutoRound's
+93.07% flip rate with a +0.275 delta is the point — high disagreement, bounded
+consequence.
+
+**Certification is about repeatability, not accuracy.** `exact_repeat: certified`
+says two runs agree bit for bit. It says nothing about whether the kernel computes
+the right thing, which is what the zero baseline (Law 1) and the reference binding
+(Laws 12 and 16) are for.

README.md

@@ -1,36 +1,46 @@
 ---
-pretty_name: gemma-4-26B-A4B-it quantization analysis
+pretty_name: "gemma-4-26B-A4B-it quantization analysis"
 tags:
-- quantization
-- kl-divergence
-- distribution-fidelity
-- evaluation
+- "quantization"
+- "kl-divergence"
+- "distribution-fidelity"
+- "evaluation"
 ---
 
 # gemma-4-26B-A4B-it quantization analysis
 
 Mean KL divergence against on-disk size
 
-Scored under the distribution-fidelity laws, version 9. Read [`LAWS.md`](LAWS.md) first: these numbers are comparable only within this artifact's token suite, geometry, and runtime identity, and not against any number produced elsewhere.
+Scored under the distribution-fidelity laws, version 12. Read [`LAWS.md`](LAWS.md) first: these numbers are comparable only within this artifact's token suite, geometry, and runtime identity, and not against any number produced elsewhere.
 
 Each candidate directory holds its one-pager (`report.md`), its raw report, its compliance receipt, and its Law 14 attribution where one was produced. `reference/` carries the reusable teacher tensors and head, so a third party can score a new candidate without loading the reference checkpoint. `checksums.txt` covers every file here.
 
-## Candidates by mean KLD
+## Candidates by paired routed-model fidelity
 
-| Candidate | Scheme | On disk | Mean KLD |
-|---|---|---|---|
-| RedHatAI/gemma-4-26B-A4B-it-FP8-dynamic | fp8_per_channel | 26.68 GiB | 0.68045608 |
-| cyankiwi/gemma-4-26B-A4B-it-AWQ-4bit | unquantized | 16.01 GiB | 1.01984639 |
-| dhruvil237/gemma-4-26B-A4B-it-W4A16 | unquantized | 14.58 GiB | 1.15699481 |
-| unsloth/gemma-4-26B-A4B-it-NVFP4 | fp8_per_channel | 15.75 GiB | 1.16528327 |
-| lcu0312/gemma-4-26B-A4B-it-AWQ-4bit | unquantized | 16.01 GiB | 1.23145061 |
-| cklaus/gemma-4-26B-A4B-it-NVFP4 | nvfp4 | 16.78 GiB | 1.57775650 |
-| RedHatAI/gemma-4-26B-A4B-it-NVFP4 | unquantized | 15.30 GiB | 1.75937795 |
-| Neural-ICE/Gemma-4-26B-A4B-it-NVFP4 | nvfp4 | 15.29 GiB | 1.81298048 |
-| bg-digitalservices/Gemma-4-26B-A4B-it-NVFP4 | nvfp4 | 15.29 GiB | 1.82145209 |
+| Candidate | Scheme | On disk | QxQ KLD | BxQ KLD | QxQ - BxQ | Natural route flips | Exact repeat |
+|---|---|---|---|---|---|---|---|
+| RedHatAI/gemma-4-26B-A4B-it-FP8-dynamic | fp8_per_channel | 26.68 GiB | 0.69415039 | 0.49783177 | +0.19631862 | 50.6828% | certified |
+| cyankiwi/gemma-4-26B-A4B-it-AWQ-4bit | int4_g32_sym | 16.01 GiB | 1.02626820 | 0.84088786 | +0.18538034 | 62.7369% | certified |
+| dhruvil237/gemma-4-26B-A4B-it-W4A16 | int4_g64_sym | 14.58 GiB | 1.15926068 | 1.01252351 | +0.14673717 | 66.7300% | certified |
+| unsloth/gemma-4-26B-A4B-it-NVFP4 | nvfp4 | 15.75 GiB | 1.16299506 | 1.00117013 | +0.16182492 | 66.4622% | certified |
+| Intel/gemma-4-26B-A4B-it-int4-AutoRound | int4_g128_sym | 14.31 GiB | 1.17816288 | 0.90308200 | +0.27508088 | 93.0727% | certified |
+| lcu0312/gemma-4-26B-A4B-it-AWQ-4bit | int4_g32_sym | 16.01 GiB | 1.23587221 | 1.08009636 | +0.15577585 | 66.1806% | certified |
+| RedHatAI/gemma-4-26B-A4B-it-NVFP4 | nvfp4 | 15.30 GiB | 1.75580190 | 1.53130858 | +0.22449332 | 74.3937% | certified |
+| Neural-ICE/Gemma-4-26B-A4B-it-NVFP4 | nvfp4 | 15.29 GiB | 1.81880328 | 1.58869884 | +0.23010445 | 74.4082% | certified |
+| bg-digitalservices/Gemma-4-26B-A4B-it-NVFP4 | nvfp4 | 15.29 GiB | 1.82545539 | 1.58674213 | +0.23871325 | 74.4254% | certified |
 
-## Fidelity against size
+## QxQ fidelity against size
 
-![mean KLD against on-disk size](kld-vs-size.png)
+![QxQ KLD against on-disk size](qxq-vs-size.png)
 
-The numbers behind the chart are in [`kld-vs-size.json`](kld-vs-size.json), because a picture is not evidence.
+## BxQ fidelity against size
+
+![BxQ KLD against on-disk size](bxq-vs-size.png)
+
+The numbers behind the charts are in [`kld-vs-size.json`](kld-vs-size.json), because a picture is not evidence.
+
+## Explicitly excluded checkpoints
+
+These pinned revisions were deliberately not scored. They are listed so a known-bad checkpoint is never silently retried or mistaken for an omitted result.
+
+- `cklaus/gemma-4-26B-A4B-it-NVFP4@54218eb0fd33da17632f0d91be841d09a7724a6c` — Scored on vLLM fdc0b57e while every other candidate scored on c798225a, so its report binds a reference capture the family no longer publishes (Law 12) and it ranks in a comparability group of one. Withdrawn rather than rescored: no measurement of this checkpoint is in question, only the provenance it carries.

baselines/self-kld.json

@@ -1,42 +1,43 @@
 {
-  "capture_manifest_sha256": "4ff0941058937d771cbe239fb31399357899dc0341f3668ea4bfb8517525aba5",
+  "capture_manifest_sha256": "305f8014578ad62130ab0ab7f2e584eb96a4f5dc330ae7fc8bc0097a35737ca2",
   "confidence_buckets": [
     {
-      "frac": 0.07572056668295066,
+      "frac": 0.06888128969223253,
       "hi": 0.25,
       "lo": 0.0,
       "mean_kld": 0.0,
-      "n": 155
+      "n": 141
     },
     {
-      "frac": 0.2525647288715193,
+      "frac": 0.26917440156326333,
       "hi": 0.5,
       "lo": 0.25,
       "mean_kld": 0.0,
-      "n": 517
+      "n": 551
     },
     {
-      "frac": 0.2530532486565706,
+      "frac": 0.23888617489008304,
       "hi": 0.75,
       "lo": 0.5,
       "mean_kld": 0.0,
-      "n": 518
+      "n": 489
     },
     {
-      "frac": 0.2325354176844162,
+      "frac": 0.23155837811431362,
       "hi": 0.95,
       "lo": 0.75,
       "mean_kld": 0.0,
-      "n": 476
+      "n": 474
     },
     {
-      "frac": 0.18612603810454323,
+      "frac": 0.19149975574010747,
       "hi": 1.0,
       "lo": 0.95,
       "mean_kld": 0.0,
-      "n": 381
+      "n": 392
     }
   ],
+  "context_length": 2048,
   "depth_buckets": [
     {
       "depth_hi": 511,
@@ -78,7 +79,7 @@
       "context_id": 0,
       "max_kld": 0.0,
       "mean_kld": 0.0,
-      "mean_ref_top1_prob": 0.6499947183667653,
+      "mean_ref_top1_prob": 0.6507112201345101,
       "median_kld": 0.0,
       "p99_kld": 0.0,
       "positions": 2047,
@@ -86,6 +87,41 @@
       "top1_agreement": 1.0
     }
   ],
+  "reference_weights_sha256": "201ebceb92c5b166cc028c3008607ba28c478c7699b1d382e7be7ede81f2ab75",
+  "runtime_binding": {
+    "compiled_extensions": {
+      "_C_stable_libtorch.abi3.so": "fbae4a92ec2861cd979eb99a3867049f88f7fcb278337be058dd4cf2ac8c29bf",
+      "_moe_C_stable_libtorch.abi3.so": "a5f41231d3a2fe2958aad06615137ab877671baea20e78b0257516ef3467664e",
+      "third_party/deep_gemm/_C.cpython-310-x86_64-linux-gnu.so": "38b87140577c2d72c2b3a6407c5f07da53ef9362e18fe988162afa8f3f91746e",
+      "third_party/deep_gemm/_C.cpython-311-x86_64-linux-gnu.so": "1f3a3430c188295cf96a5ac977ec72258bccb525daaa7b0ad6642bf3a995d3da",
+      "third_party/deep_gemm/_C.cpython-312-x86_64-linux-gnu.so": "73824dc1312e98cf277ae6bc017becf7e963d5b0bfab81e0c6d475d9b769c9e7",
+      "third_party/deep_gemm/_C.cpython-313-x86_64-linux-gnu.so": "bd0124b39fb500378764bf0d1adb5fe30ea876e66ed65971d8f668de4a8dbe83",
+      "third_party/deep_gemm/_C.cpython-314-x86_64-linux-gnu.so": "0f9817dfe8a619dde56456c6e7a5a9d914fede0fb145bc48aad134e19c77fff0"
+    },
+    "compiled_extensions_sha256": "f2fbc7537b0f01f65341030cfa90741c2ed7ad46c88b0faa3dd0b46d311ccdc8",
+    "cuda": "13.2",
+    "determinism": {
+      "CUBLAS_WORKSPACE_CONFIG": ":4096:8",
+      "NCCL_DETERMINISTIC": "1",
+      "VLLM_BATCH_INVARIANT": "1",
+      "VLLM_MOE_USE_DEEP_GEMM": "0"
+    },
+    "driver": "580.173.02",
+    "flashinfer": "0.6.17",
+    "gpu_names": [
+      "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
+      "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
+      "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
+      "NVIDIA RTX PRO 6000 Blackwell Workstation Edition"
+    ],
+    "nccl": "2.29.7",
+    "platform": "Linux-6.8.0-137-generic-x86_64-with-glibc2.39",
+    "python": "3.12.3",
+    "torch": "2.13.0+cu132",
+    "vllm_commit": "c798225a7f197994ed502068fc6ac1ba8a9706f3",
+    "vllm_dirty_digest": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+    "vllm_tree_dirty": false
+  },
   "score_from": 0,
   "storage": "hidden",
   "student_lm_head": {
@@ -185,12 +221,12 @@
     }
   },
   "timings": {
-    "capture_forward": 1.645566038088873,
-    "capture_write": 0.5230729130562395,
-    "replay_probe": 3.432808293029666,
-    "score_forward": 0.48364015319384634,
-    "student_load": 19.434812565799803,
-    "teacher_load": 30.872963302070275
+    "capture_forward": 1.5388168222270906,
+    "capture_write": 0.5069590341299772,
+    "replay_probe": 3.009788952767849,
+    "score_forward": 0.47462915861979127,
+    "student_load": 19.50744666904211,
+    "teacher_load": 22.79794552316889
   },
   "top1_agreement": 1.0,
   "topk_agreement": {
@@ -201,5 +237,5 @@
     "5": 1.0
   },
   "unique_tokens": 2048,
-  "student_weights_sha256": "7d8dba482b208253b9fa0f77d0943691b09e86e0677bf613d177c037d2d506a2"
+  "student_weights_sha256": "201ebceb92c5b166cc028c3008607ba28c478c7699b1d382e7be7ede81f2ab75"
 }

environment/gpu-clocks.txt

@@ -1,7 +1,7 @@
 ### nvidia-smi --query-gpu=index,name,serial,uuid,driver_version,vbios_version,pstate,clocks.sm,clocks.max.sm,power.limit,persistence_mode,ecc.mode.current --format=csv
 
 index, name, serial, uuid, driver_version, vbios_version, pstate, clocks.current.sm [MHz], clocks.max.sm [MHz], power.limit [W], persistence_mode, ecc.mode.current
-0, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 1794225038968, GPU-41235b51-d6ab-afe6-c393-e1c377b81299, 580.173.02, 98.02.81.00.07, P0, 2475 MHz, 4087 MHz, 575.00 W, Disabled, Disabled
+0, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 1794225038968, GPU-41235b51-d6ab-afe6-c393-e1c377b81299, 580.173.02, 98.02.81.00.07, P8, 1170 MHz, 4087 MHz, 575.00 W, Disabled, Disabled
 1, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 1793025105275, GPU-547cb69a-4bcf-5f0a-c154-5397b1dc5a78, 580.173.02, 98.02.81.00.07, P8, 1177 MHz, 4087 MHz, 575.00 W, Disabled, Disabled
 2, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 1792925028706, GPU-0f14730c-9c8d-aa20-9478-324d31766353, 580.173.02, 98.02.81.00.07, P8, 1177 MHz, 4080 MHz, 575.00 W, Disabled, Disabled
 3, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 1794125045055, GPU-2aea821a-9e57-5a64-9471-83306dddcd97, 580.173.02, 98.02.81.00.07, P8, 1177 MHz, 4087 MHz, 575.00 W, Disabled, Disabled

environment/gpu-smi-query.txt

@@ -3,7 +3,7 @@
 
 ==============NVSMI LOG==============
 
-Timestamp                                              : Wed Sep  2 20:09:29 2026
+Timestamp                                              : Tue Sep  8 01:43:22 2026
 Driver Version                                         : 580.173.02
 CUDA Version                                           : 13.0
 
@@ -49,8 +49,8 @@ GPU 00000000:01:00.0
         ECC Object                                     : 7.16
         Power Management Object                        : N/A
     Inforom BBX Object Flush
-        Latest Timestamp                               : 2026/09/02 19:35:56.352
-        Latest Duration                                : 58298 us
+        Latest Timestamp                               : 2026/09/07 20:07:16.694
+        Latest Duration                                : 30306 us
     GPU Operation Mode
         Current                                        : N/A
         Pending                                        : N/A
@@ -75,8 +75,8 @@ GPU 00000000:01:00.0
         GPU Link Info
             PCIe Generation
                 Max                                    : 5
-                Current                                : 5
-                Device Current                         : 5
+                Current                                : 1
+                Device Current                         : 1
                 Device Max                             : 5
                 Host Max                               : 5
             Link Width
@@ -85,14 +85,14 @@ GPU 00000000:01:00.0
         Bridge Chip
             Type                                       : N/A
             Firmware                                   : N/A
-        Replays Since Reset                            : 38009
+        Replays Since Reset                            : 41090
         Replay Number Rollovers                        : 0
-        Tx Throughput                                  : 837 KB/s
-        Rx Throughput                                  : 510 KB/s
+        Tx Throughput                                  : 2259 KB/s
+        Rx Throughput                                  : 1494 KB/s
         Atomic Caps Outbound                           : N/A
         Atomic Caps Inbound                            : FETCHADD_32 FETCHADD_64 SWAP_32 SWAP_64 CAS_32 CAS_64 
     Fan Speed                                          : 30 %
-    Performance State                                  : P1
+    Performance State                                  : P8
     Clocks Event Reasons
         Idle                                           : Not Active
         Applications Clocks Setting                    : Not Active
@@ -104,7 +104,7 @@ GPU 00000000:01:00.0
         SW Thermal Slowdown                            : Not Active
         Display Clock Setting                          : Not Active
     Clocks Event Reasons Counters
-        SW Power Capping                               : 2614908053 us
+        SW Power Capping                               : 2615809600 us
         Sync Boost                                     : 0 us
         SW Thermal Slowdown                            : 0 us
         HW Thermal Slowdown                            : 0 us
@@ -113,12 +113,12 @@ GPU 00000000:01:00.0
     FB Memory Usage
         Total                                          : 97887 MiB
         Reserved                                       : 640 MiB
-        Used                                           : 2 MiB
-        Free                                           : 97247 MiB
+        Used                                           : 4 MiB
+        Free                                           : 97244 MiB
     BAR1 Memory Usage
         Total                                          : 256 MiB
-        Used                                           : 2 MiB
-        Free                                           : 254 MiB
+        Used                                           : 3 MiB
+        Free                                           : 253 MiB
     Conf Compute Protected Memory Usage
         Total                                          : 0 MiB
         Used                                           : 0 MiB
@@ -186,8 +186,8 @@ GPU 00000000:01:00.0
             Low                                        : 0 bank(s)
             None                                       : 0 bank(s)
     Temperature
-        GPU Current Temp                               : 33 C
-        GPU T.Limit Temp                               : 60 C
+        GPU Current Temp                               : 30 C
+        GPU T.Limit Temp                               : 63 C
         GPU Shutdown T.Limit Temp                      : -5 C
         GPU Slowdown T.Limit Temp                      : -2 C
         GPU Max Operating T.Limit Temp                 : 0 C
@@ -195,8 +195,8 @@ GPU 00000000:01:00.0
         Memory Current Temp                            : N/A
         Memory Max Operating T.Limit Temp              : N/A
     GPU Power Readings
-        Average Power Draw                             : 71.92 W
-        Instantaneous Power Draw                       : 73.61 W
+        Average Power Draw                             : 19.10 W
+        Instantaneous Power Draw                       : 19.86 W
         Current Power Limit                            : 575.00 W
         Requested Power Limit                          : 575.00 W
         Default Power Limit                            : 600.00 W
@@ -218,10 +218,10 @@ GPU 00000000:01:00.0
         Requested Profiles                             : N/A
         Enforced Profiles                              : N/A
     Clocks
-        Graphics                                       : 2475 MHz
-        SM                                             : 2475 MHz
-        Memory                                         : 14865 MHz
-        Video                                          : 1950 MHz
+        Graphics                                       : 1170 MHz
+        SM                                             : 1170 MHz
+        Memory                                         : 405 MHz
+        Video                                          : 600 MHz
     Applications Clocks
         Graphics                                       : 2617 MHz
         Memory                                         : 14001 MHz
@@ -298,8 +298,8 @@ GPU 00000000:21:00.0
         ECC Object                                     : 7.16
         Power Management Object                        : N/A
     Inforom BBX Object Flush
-        Latest Timestamp                               : 2026/09/02 19:36:18.086
-        Latest Duration                                : 64784 us
+        Latest Timestamp                               : 2026/09/07 20:06:52.409
+        Latest Duration                                : 68770 us
     GPU Operation Mode
         Current                                        : N/A
         Pending                                        : N/A
@@ -334,10 +334,10 @@ GPU 00000000:21:00.0
         Bridge Chip
             Type                                       : N/A
             Firmware                                   : N/A
-        Replays Since Reset                            : 146
+        Replays Since Reset                            : 155
         Replay Number Rollovers                        : 0
-        Tx Throughput                                  : 1471 KB/s
-        Rx Throughput                                  : 483 KB/s
+        Tx Throughput                                  : 2265 KB/s
+        Rx Throughput                                  : 1672 KB/s
         Atomic Caps Outbound                           : N/A
         Atomic Caps Inbound                            : FETCHADD_32 FETCHADD_64 SWAP_32 SWAP_64 CAS_32 CAS_64 
     Fan Speed                                          : 30 %
@@ -362,12 +362,12 @@ GPU 00000000:21:00.0
     FB Memory Usage
         Total                                          : 97887 MiB
         Reserved                                       : 640 MiB
-        Used                                           : 2 MiB
-        Free                                           : 97247 MiB
+        Used                                           : 4 MiB
+        Free                                           : 97244 MiB
     BAR1 Memory Usage
         Total                                          : 256 MiB
-        Used                                           : 2 MiB
-        Free                                           : 254 MiB
+        Used                                           : 3 MiB
+        Free                                           : 253 MiB
     Conf Compute Protected Memory Usage
         Total                                          : 0 MiB
         Used                                           : 0 MiB
@@ -444,8 +444,8 @@ GPU 00000000:21:00.0
         Memory Current Temp                            : N/A
         Memory Max Operating T.Limit Temp              : N/A
     GPU Power Readings
-        Average Power Draw                             : 9.38 W
-        Instantaneous Power Draw                       : 13.39 W
+        Average Power Draw                             : 14.75 W
+        Instantaneous Power Draw                       : 16.58 W
         Current Power Limit                            : 575.00 W
         Requested Power Limit                          : 575.00 W
         Default Power Limit                            : 600.00 W
@@ -547,8 +547,8 @@ GPU 00000000:41:00.0
         ECC Object                                     : 7.16
         Power Management Object                        : N/A
     Inforom BBX Object Flush
-        Latest Timestamp                               : 2026/09/02 19:37:21.785
-        Latest Duration                                : 68842 us
+        Latest Timestamp                               : 2026/09/07 20:07:00.910
+        Latest Duration                                : 66539 us
     GPU Operation Mode
         Current                                        : N/A
         Pending                                        : N/A
@@ -585,8 +585,8 @@ GPU 00000000:41:00.0
             Firmware                                   : N/A
         Replays Since Reset                            : 3
         Replay Number Rollovers                        : 0
-        Tx Throughput                                  : 892 KB/s
-        Rx Throughput                                  : 533 KB/s
+        Tx Throughput                                  : 2213 KB/s
+        Rx Throughput                                  : 1507 KB/s
         Atomic Caps Outbound                           : N/A
         Atomic Caps Inbound                            : FETCHADD_32 FETCHADD_64 SWAP_32 SWAP_64 CAS_32 CAS_64 
     Fan Speed                                          : 30 %
@@ -611,12 +611,12 @@ GPU 00000000:41:00.0
     FB Memory Usage
         Total                                          : 97887 MiB
         Reserved                                       : 640 MiB
-        Used                                           : 2 MiB
-        Free                                           : 97247 MiB
+        Used                                           : 4 MiB
+        Free                                           : 97244 MiB
     BAR1 Memory Usage
         Total                                          : 256 MiB
-        Used                                           : 2 MiB
-        Free                                           : 254 MiB
+        Used                                           : 3 MiB
+        Free                                           : 253 MiB
     Conf Compute Protected Memory Usage
         Total                                          : 0 MiB
         Used                                           : 0 MiB
@@ -693,8 +693,8 @@ GPU 00000000:41:00.0
         Memory Current Temp                            : N/A
         Memory Max Operating T.Limit Temp              : N/A
     GPU Power Readings
-        Average Power Draw                             : 16.60 W
-        Instantaneous Power Draw                       : 20.55 W
+        Average Power Draw                             : 19.84 W
+        Instantaneous Power Draw                       : 20.94 W
         Current Power Limit                            : 575.00 W
         Requested Power Limit                          : 575.00 W
         Default Power Limit                            : 600.00 W
@@ -796,8 +796,8 @@ GPU 00000000:C1:00.0
         ECC Object                                     : 7.16
         Power Management Object                        : N/A
     Inforom BBX Object Flush
-        Latest Timestamp                               : 2026/09/01 21:13:49.704
-        Latest Duration                                : 30647 us
+        Latest Timestamp                               : 2026/09/07 21:19:49.087
+        Latest Duration                                : 66464 us
     GPU Operation Mode
         Current                                        : N/A
         Pending                                        : N/A
@@ -834,8 +834,8 @@ GPU 00000000:C1:00.0
             Firmware                                   : N/A
         Replays Since Reset                            : 0
         Replay Number Rollovers                        : 0
-        Tx Throughput                                  : 670 KB/s
-        Rx Throughput                                  : 438 KB/s
+        Tx Throughput                                  : 1957 KB/s
+        Rx Throughput                                  : 1494 KB/s
         Atomic Caps Outbound                           : N/A
         Atomic Caps Inbound                            : FETCHADD_32 FETCHADD_64 SWAP_32 SWAP_64 CAS_32 CAS_64 
     Fan Speed                                          : 30 %
@@ -860,12 +860,12 @@ GPU 00000000:C1:00.0
     FB Memory Usage
         Total                                          : 97887 MiB
         Reserved                                       : 640 MiB
-        Used                                           : 2 MiB
-        Free                                           : 97247 MiB
+        Used                                           : 4 MiB
+        Free                                           : 97244 MiB
     BAR1 Memory Usage
         Total                                          : 256 MiB
-        Used                                           : 2 MiB
-        Free                                           : 254 MiB
+        Used                                           : 3 MiB
+        Free                                           : 253 MiB
     Conf Compute Protected Memory Usage
         Total                                          : 0 MiB
         Used                                           : 0 MiB
@@ -942,8 +942,8 @@ GPU 00000000:C1:00.0
         Memory Current Temp                            : N/A
         Memory Max Operating T.Limit Temp              : N/A
     GPU Power Readings
-        Average Power Draw                             : 17.72 W
-        Instantaneous Power Draw                       : 20.30 W
+        Average Power Draw                             : 18.77 W
+        Instantaneous Power Draw                       : 17.95 W
         Current Power Limit                            : 575.00 W
         Requested Power Limit                          : 575.00 W
         Default Power Limit                            : 600.00 W

environment/gpu-smi.txt

@@ -1,6 +1,6 @@
 ### nvidia-smi
 
-Wed Sep  2 20:09:29 2026       
+Tue Sep  8 01:43:22 2026       
 +-----------------------------------------------------------------------------------------+
 | NVIDIA-SMI 580.173.02             Driver Version: 580.173.02     CUDA Version: 13.0     |
 +-----------------------------------------+------------------------+----------------------+
@@ -9,19 +9,19 @@ Wed Sep  2 20:09:29 2026
 |                                         |                        |               MIG M. |
 |=========================================+========================+======================|
 |   0  NVIDIA RTX PRO 6000 Blac...    Off |   00000000:01:00.0 Off |                  Off |
-| 30%   33C    P1             71W /  575W |       2MiB /  97887MiB |      0%      Default |
+| 30%   30C    P8             18W /  575W |       4MiB /  97887MiB |      0%      Default |
 |                                         |                        |                  N/A |
 +-----------------------------------------+------------------------+----------------------+
 |   1  NVIDIA RTX PRO 6000 Blac...    Off |   00000000:21:00.0 Off |                  Off |
-| 30%   29C    P8              7W /  575W |       2MiB /  97887MiB |      0%      Default |
+| 30%   29C    P8             10W /  575W |       4MiB /  97887MiB |      0%      Default |
 |                                         |                        |                  N/A |
 +-----------------------------------------+------------------------+----------------------+
 |   2  NVIDIA RTX PRO 6000 Blac...    Off |   00000000:41:00.0 Off |                  Off |
-| 30%   30C    P8             15W /  575W |       2MiB /  97887MiB |      0%      Default |
+| 30%   30C    P8             17W /  575W |       4MiB /  97887MiB |      0%      Default |
 |                                         |                        |                  N/A |
 +-----------------------------------------+------------------------+----------------------+
 |   3  NVIDIA RTX PRO 6000 Blac...    Off |   00000000:C1:00.0 Off |                  Off |
-| 30%   29C    P8             15W /  575W |       2MiB /  97887MiB |      0%      Default |
+| 30%   29C    P8             16W /  575W |       4MiB /  97887MiB |      0%      Default |
 |                                         |                        |                  N/A |
 +-----------------------------------------+------------------------+----------------------+
 

environment/host-cpu.txt

@@ -15,7 +15,7 @@ Core(s) per socket:                      24
 Socket(s):                               1
 Stepping:                                1
 Frequency boost:                         enabled
-CPU(s) scaling MHz:                      56%
+CPU(s) scaling MHz:                      51%
 CPU max MHz:                             4816.6992
 CPU min MHz:                             1500.0000
 BogoMIPS:                                8200.02

environment/host-memory.txt

@@ -1,7 +1,7 @@
 ### free -g
 
                total        used        free      shared  buff/cache   available
-Mem:             755          16         138           0         605         739
+Mem:             755          15         163           0         581         739
 Swap:              7           0           7
 
 ### exit=0

environment/host-mounts.txt

@@ -1,13 +1,13 @@
 ### df -h
 
 Filesystem                                  Size  Used Avail Use% Mounted on
-tmpfs                                        76G  5.6M   76G   1% /run
+tmpfs                                        76G  5.5M   76G   1% /run
 efivarfs                                    256K  125K  127K  50% /sys/firmware/efi/efivars
-/dev/nvme0n1p2                              3.6T  3.1T  359G  90% /
+/dev/nvme0n1p2                              3.6T  3.1T  354G  90% /
 tmpfs                                       378G     0  378G   0% /dev/shm
 tmpfs                                       5.0M     0  5.0M   0% /run/lock
 /dev/nvme0n1p1                              1.1G  6.2M  1.1G   1% /boot/efi
-/dev/nvme1n1p1                              3.6T  1.9T  1.6T  54% /media/fmodels2
+/dev/nvme1n1p1                              3.6T  1.8T  1.7T  52% /media/fmodels2
 tmpfs                                        76G   20K   76G   1% /run/user/1000
 10.9.99.242:/mnt/GANAS02-T02S01/Ai-Storage   34T   21T   13T  62% /media/netmodels
 

environment/models/gemma-4-26B-A4B-it-W4A16.txt

@@ -1,3 +1,432 @@
 ### /media/fmodels2/dhruvil237/gemma-4-26B-A4B-it-W4A16
 
-not a directory
+## file listing (name size)
+chat_template.jinja 12045
+config.json 19270
+generation_config.json 203
+.gitattributes 1570
+processor_config.json 1689
+README.md 1284
+recipe.yaml 934
+tokenizer_config.json 2712
+tokenizer.json 32169626
+
+## listing sha256
+0224dd5c4bbfe0b881c212e109312a29dcc818128a4f4780c17f79ecb798c9a6  -
+
+## total bytes / file count
+7945391992	/media/fmodels2/dhruvil237/gemma-4-26B-A4B-it-W4A16
+32
+
+## config and tokenizer sha256
+773a83885006e4b425036ea214b15889fff5a04482063c4a2d9a3074f72df9bb  /media/fmodels2/dhruvil237/gemma-4-26B-A4B-it-W4A16/config.json
+5439541d3bf0ba9ade7c1122dc14703f85f32a520150351b344b14f48c574cc2  /media/fmodels2/dhruvil237/gemma-4-26B-A4B-it-W4A16/generation_config.json
+db33eb89116392fac778165935900c87495a0b357d43fea009c77be7d96134d0  /media/fmodels2/dhruvil237/gemma-4-26B-A4B-it-W4A16/tokenizer_config.json
+cc8d3a0ce36466ccc1278bf987df5f71db1719b9ca6b4118264f45cb627bfe0f  /media/fmodels2/dhruvil237/gemma-4-26B-A4B-it-W4A16/tokenizer.json
+2dfbfc7d538912f4ea11d29d85b4e25d7bc26386e53f57529f1d707c28b5828c  /media/fmodels2/dhruvil237/gemma-4-26B-A4B-it-W4A16/chat_template.jinja
+
+## config.json
+{
+  "architectures": [
+    "Gemma4ForConditionalGeneration"
+  ],
+  "audio_config": null,
+  "audio_token_id": 258881,
+  "boa_token_id": 256000,
+  "boi_token_id": 255999,
+  "dtype": "bfloat16",
+  "eoa_token_id": 258883,
+  "eoa_token_index": 258883,
+  "eoi_token_id": 258882,
+  "eos_token_id": [
+    1,
+    106
+  ],
+  "image_token_id": 258880,
+  "initializer_range": 0.02,
+  "model_type": "gemma4",
+  "quantization_config": {
+    "config_groups": {
+      "group_0": {
+        "format": "pack-quantized",
+        "input_activations": null,
+        "output_activations": null,
+        "targets": [
+          "Linear"
+        ],
+        "weights": {
+          "actorder": "static",
+          "block_structure": null,
+          "dynamic": false,
+          "group_size": 64,
+          "num_bits": 4,
+          "observer": "memoryless_minmax",
+          "observer_kwargs": {},
+          "scale_dtype": null,
+          "strategy": "group",
+          "symmetric": true,
+          "type": "int",
+          "zp_dtype": null
+        }
+      }
+    },
+    "format": "pack-quantized",
+    "global_compression_ratio": null,
+    "ignore": [
+      "model.language_model.layers.0.router.proj",
+      "model.language_model.layers.1.router.proj",
+      "model.language_model.layers.2.router.proj",
+      "model.language_model.layers.3.router.proj",
+      "model.language_model.layers.4.router.proj",
+      "model.language_model.layers.5.router.proj",
+      "model.language_model.layers.6.router.proj",
+      "model.language_model.layers.7.router.proj",
+      "model.language_model.layers.8.router.proj",
+      "model.language_model.layers.9.router.proj",
+      "model.language_model.layers.10.router.proj",
+      "model.language_model.layers.11.router.proj",
+      "model.language_model.layers.12.router.proj",
+      "model.language_model.layers.13.router.proj",
+      "model.language_model.layers.14.router.proj",
+      "model.language_model.layers.15.router.proj",
+      "model.language_model.layers.16.router.proj",
+      "model.language_model.layers.17.router.proj",
+      "model.language_model.layers.18.router.proj",
+      "model.language_model.layers.19.router.proj",
+      "model.language_model.layers.20.router.proj",
+      "model.language_model.layers.21.router.proj",
+      "model.language_model.layers.22.router.proj",
+      "model.language_model.layers.23.router.proj",
+      "model.language_model.layers.24.router.proj",
+      "model.language_model.layers.25.router.proj",
+      "model.language_model.layers.26.router.proj",
+      "model.language_model.layers.27.router.proj",
+      "model.language_model.layers.28.router.proj",
+      "model.language_model.layers.29.router.proj",
+      "model.vision_tower.patch_embedder.input_proj",
+      "model.vision_tower.encoder.layers.0.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.0.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.0.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.0.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.0.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.0.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.0.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.1.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.1.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.1.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.1.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.1.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.1.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.1.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.2.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.2.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.2.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.2.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.2.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.2.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.2.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.3.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.3.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.3.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.3.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.3.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.3.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.3.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.4.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.4.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.4.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.4.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.4.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.4.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.4.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.5.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.5.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.5.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.5.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.5.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.5.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.5.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.6.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.6.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.6.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.6.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.6.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.6.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.6.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.7.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.7.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.7.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.7.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.7.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.7.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.7.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.8.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.8.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.8.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.8.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.8.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.8.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.8.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.9.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.9.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.9.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.9.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.9.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.9.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.9.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.10.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.10.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.10.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.10.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.10.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.10.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.10.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.11.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.11.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.11.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.11.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.11.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.11.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.11.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.12.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.12.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.12.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.12.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.12.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.12.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.12.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.13.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.13.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.13.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.13.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.13.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.13.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.13.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.14.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.14.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.14.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.14.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.14.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.14.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.14.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.15.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.15.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.15.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.15.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.15.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.15.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.15.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.16.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.16.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.16.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.16.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.16.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.16.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.16.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.17.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.17.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.17.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.17.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.17.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.17.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.17.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.18.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.18.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.18.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.18.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.18.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.18.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.18.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.19.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.19.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.19.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.19.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.19.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.19.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.19.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.20.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.20.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.20.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.20.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.20.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.20.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.20.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.21.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.21.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.21.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.21.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.21.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.21.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.21.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.22.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.22.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.22.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.22.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.22.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.22.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.22.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.23.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.23.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.23.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.23.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.23.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.23.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.23.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.24.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.24.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.24.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.24.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.24.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.24.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.24.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.25.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.25.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.25.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.25.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.25.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.25.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.25.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.26.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.26.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.26.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.26.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.26.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.26.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.26.mlp.down_proj.linear",
+      "model.embed_vision.embedding_projection",
+      "lm_head"
+    ],
+    "kv_cache_scheme": null,
+    "quant_method": "compressed-tensors",
+    "quantization_status": "compressed",
+    "sparsity_config": {},
+    "transform_config": {},
+    "version": "0.15.0"
+  },
+  "text_config": {
+    "attention_bias": false,
+    "attention_dropout": 0.0,
+    "attention_k_eq_v": true,
+    "bos_token_id": 2,
+    "dtype": "bfloat16",
+    "enable_moe_block": true,
+    "eos_token_id": 1,
+    "final_logit_softcapping": 30.0,
+    "global_head_dim": 512,
+    "head_dim": 256,
+    "hidden_activation": "gelu_pytorch_tanh",
+    "hidden_size": 2816,
+    "hidden_size_per_layer_input": 0,
+    "initializer_range": 0.02,
+    "intermediate_size": 2112,
+    "layer_types": [
+      "sliding_attention",
+      "sliding_attention",
+      "sliding_attention",
+      "sliding_attention",
+      "sliding_attention",
+      "full_attention",
+      "sliding_attention",
+      "sliding_attention",
+      "sliding_attention",
+      "sliding_attention",
+      "sliding_attention",
+      "full_attention",
+      "sliding_attention",
+      "sliding_attention",
+      "sliding_attention",
+      "sliding_attention",
+      "sliding_attention",
+      "full_attention",
+      "sliding_attention",
+      "sliding_attention",
+      "sliding_attention",
+      "sliding_attention",
+      "sliding_attention",
+      "full_attention",
+      "sliding_attention",
+      "sliding_attention",
+      "sliding_attention",
+      "sliding_attention",
+      "sliding_attention",
+      "full_attention"
+    ],
+    "max_position_embeddings": 262144,
+    "model_type": "gemma4_text",
+    "moe_intermediate_size": 704,
+    "num_attention_heads": 16,
+    "num_experts": 128,
+    "num_global_key_value_heads": 2,
+    "num_hidden_layers": 30,
+    "num_key_value_heads": 8,
+    "num_kv_shared_layers": 0,
+    "pad_token_id": 0,
+    "rms_norm_eps": 1e-06,
+    "rope_parameters": {
+      "full_attention": {
+        "partial_rotary_factor": 0.25,
+        "rope_theta": 1000000.0,
+        "rope_type": "proportional"
+      },
+      "sliding_attention": {
+        "rope_theta": 10000.0,
+        "rope_type": "default"
+      }
+    },
+    "sliding_window": 1024,
+    "tie_word_embeddings": true,
+    "top_k_experts": 8,
+    "use_bidirectional_attention": "vision",
+    "use_cache": true,
+    "use_double_wide_mlp": false,
+    "vocab_size": 262144,
+    "vocab_size_per_layer_input": 262144
+  },
+  "tie_word_embeddings": true,
+  "transformers_version": "5.5.0",
+  "video_token_id": 258884,
+  "vision_config": {
+    "_name_or_path": "",
+    "architectures": null,
+    "attention_bias": false,
+    "attention_dropout": 0.0,
+    "chunk_size_feed_forward": 0,
+    "default_output_length": 280,
+    "dtype": "bfloat16",
+    "global_head_dim": 72,
+    "head_dim": 72,
+    "hidden_activation": "gelu_pytorch_tanh",
+    "hidden_size": 1152,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "initializer_range": 0.02,
+    "intermediate_size": 4304,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "max_position_embeddings": 131072,
+    "model_type": "gemma4_vision",
+    "num_attention_heads": 16,
+    "num_hidden_layers": 27,
+    "num_key_value_heads": 16,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "patch_size": 16,
+    "pooling_kernel_size": 3,
+    "position_embedding_size": 10240,
+    "problem_type": null,
+    "return_dict": true,
+    "rms_norm_eps": 1e-06,
+    "rope_parameters": {
+      "rope_theta": 100.0,
+      "rope_type": "default"
+    },
+    "standardize": true,
+    "use_clipped_linears": false
+  },
+  "vision_soft_tokens_per_image": 280
+}

environment/models/gemma-4-26B-A4B-it-int4-AutoRound.txt

@@ -1,197 +1,3 @@
 ### /media/fmodels2/Intel/gemma-4-26B-A4B-it-int4-AutoRound
 
-## file listing (name size)
-chat_template.jinja 12045
-config.json 4096
-generation_config.json 203
-.gitattributes 1570
-inspect.json 564
-model-00001-of-00008.safetensors 2146520500
-model-00002-of-00008.safetensors 2148085850
-model-00003-of-00008.safetensors 2148148186
-model-00004-of-00008.safetensors 2148151698
-model-00005-of-00008.safetensors 2148148178
-model-00006-of-00008.safetensors 2006177808
-model-00007-of-00008.safetensors 2141640496
-model-00008-of-00008.safetensors 480398920
-model.safetensors.index.json 3649773
-preprocessor_config.json 403
-processor_config.json 1689
-quantization_config.json 241
-README.md 3591
-tokenizer_config.json 2685
-tokenizer.json 32169626
-
-## listing sha256
-0515d9e276f57893cd780b2a1e6d478b5127dbb6745b41cd266167f4884c7a68  -
-
-## total bytes / file count
-15403124403	/media/fmodels2/Intel/gemma-4-26B-A4B-it-int4-AutoRound
-61
-
-## config and tokenizer sha256
-adb3f5dd240dfaaf5645f40bf69e9ea61ccd1c897eee6db7902ec66ec3ccc2ce  /media/fmodels2/Intel/gemma-4-26B-A4B-it-int4-AutoRound/config.json
-5439541d3bf0ba9ade7c1122dc14703f85f32a520150351b344b14f48c574cc2  /media/fmodels2/Intel/gemma-4-26B-A4B-it-int4-AutoRound/generation_config.json
-b9686f02a0bb2d5de25e5bed7949d5858109e62cbb1111ce4cea731885e779e9  /media/fmodels2/Intel/gemma-4-26B-A4B-it-int4-AutoRound/tokenizer_config.json
-cc8d3a0ce36466ccc1278bf987df5f71db1719b9ca6b4118264f45cb627bfe0f  /media/fmodels2/Intel/gemma-4-26B-A4B-it-int4-AutoRound/tokenizer.json
-12476d626b89e273a9b2c6a493a6319ab02427ebfcd3635d92b29584b27c7249  /media/fmodels2/Intel/gemma-4-26B-A4B-it-int4-AutoRound/preprocessor_config.json
-2dfbfc7d538912f4ea11d29d85b4e25d7bc26386e53f57529f1d707c28b5828c  /media/fmodels2/Intel/gemma-4-26B-A4B-it-int4-AutoRound/chat_template.jinja
-0f4b4ee8b0c93d6d8eec151e28b07d1862112b940613cbbdc5dc69823c0af298  /media/fmodels2/Intel/gemma-4-26B-A4B-it-int4-AutoRound/model.safetensors.index.json
-
-## config.json
-{
-  "architectures": [
-    "Gemma4ForConditionalGeneration"
-  ],
-  "audio_config": null,
-  "audio_token_id": 258881,
-  "boa_token_id": 256000,
-  "boi_token_id": 255999,
-  "dtype": "bfloat16",
-  "eoa_token_id": 258883,
-  "eoa_token_index": 258883,
-  "eoi_token_id": 258882,
-  "eos_token_id": [
-    1,
-    106
-  ],
-  "image_token_id": 258880,
-  "initializer_range": 0.02,
-  "model_type": "gemma4",
-  "quantization_config": {
-    "autoround_version": "0.13.0",
-    "bits": 4,
-    "block_name_to_quantize": "model.language_model.layers",
-    "data_type": "int",
-    "group_size": 128,
-    "packing_format": "auto_round:auto_gptq",
-    "quant_method": "auto-round",
-    "sym": true
-  },
-  "text_config": {
-    "attention_bias": false,
-    "attention_dropout": 0.0,
-    "attention_k_eq_v": true,
-    "bos_token_id": 2,
-    "dtype": "bfloat16",
-    "enable_moe_block": true,
-    "eos_token_id": 1,
-    "final_logit_softcapping": 30.0,
-    "global_head_dim": 512,
-    "head_dim": 256,
-    "hidden_activation": "gelu_pytorch_tanh",
-    "hidden_size": 2816,
-    "hidden_size_per_layer_input": 0,
-    "initializer_range": 0.02,
-    "intermediate_size": 2112,
-    "layer_types": [
-      "sliding_attention",
-      "sliding_attention",
-      "sliding_attention",
-      "sliding_attention",
-      "sliding_attention",
-      "full_attention",
-      "sliding_attention",
-      "sliding_attention",
-      "sliding_attention",
-      "sliding_attention",
-      "sliding_attention",
-      "full_attention",
-      "sliding_attention",
-      "sliding_attention",
-      "sliding_attention",
-      "sliding_attention",
-      "sliding_attention",
-      "full_attention",
-      "sliding_attention",
-      "sliding_attention",
-      "sliding_attention",
-      "sliding_attention",
-      "sliding_attention",
-      "full_attention",
-      "sliding_attention",
-      "sliding_attention",
-      "sliding_attention",
-      "sliding_attention",
-      "sliding_attention",
-      "full_attention"
-    ],
-    "max_position_embeddings": 262144,
-    "model_type": "gemma4_text",
-    "moe_intermediate_size": 704,
-    "num_attention_heads": 16,
-    "num_experts": 128,
-    "num_global_key_value_heads": 2,
-    "num_hidden_layers": 30,
-    "num_key_value_heads": 8,
-    "num_kv_shared_layers": 0,
-    "pad_token_id": 0,
-    "rms_norm_eps": 1e-06,
-    "rope_parameters": {
-      "full_attention": {
-        "partial_rotary_factor": 0.25,
-        "rope_theta": 1000000.0,
-        "rope_type": "proportional"
-      },
-      "sliding_attention": {
-        "rope_theta": 10000.0,
-        "rope_type": "default"
-      }
-    },
-    "sliding_window": 1024,
-    "tie_word_embeddings": true,
-    "top_k_experts": 8,
-    "use_bidirectional_attention": "vision",
-    "use_cache": true,
-    "use_double_wide_mlp": false,
-    "vocab_size": 262144,
-    "vocab_size_per_layer_input": 262144
-  },
-  "tie_word_embeddings": true,
-  "transformers_version": "5.5.0",
-  "video_token_id": 258884,
-  "vision_config": {
-    "_name_or_path": "",
-    "architectures": null,
-    "attention_bias": false,
-    "attention_dropout": 0.0,
-    "chunk_size_feed_forward": 0,
-    "default_output_length": 280,
-    "dtype": "bfloat16",
-    "global_head_dim": 72,
-    "head_dim": 72,
-    "hidden_activation": "gelu_pytorch_tanh",
-    "hidden_size": 1152,
-    "id2label": {
-      "0": "LABEL_0",
-      "1": "LABEL_1"
-    },
-    "initializer_range": 0.02,
-    "intermediate_size": 4304,
-    "is_encoder_decoder": false,
-    "label2id": {
-      "LABEL_0": 0,
-      "LABEL_1": 1
-    },
-    "max_position_embeddings": 131072,
-    "model_type": "gemma4_vision",
-    "num_attention_heads": 16,
-    "num_hidden_layers": 27,
-    "num_key_value_heads": 16,
-    "output_attentions": false,
-    "output_hidden_states": false,
-    "patch_size": 16,
-    "pooling_kernel_size": 3,
-    "position_embedding_size": 10240,
-    "problem_type": null,
-    "return_dict": true,
-    "rms_norm_eps": 1e-06,
-    "rope_parameters": {
-      "rope_theta": 100.0,
-      "rope_type": "default"
-    },
-    "standardize": true,
-    "use_clipped_linears": false
-  },
-  "vision_soft_tokens_per_image": 280
-}
+not a directory

environment/pip-freeze.txt

@@ -20,6 +20,7 @@ charset-normalizer==3.5.1
 click==8.5.0
 cloudpickle==3.1.2
 compressed-tensors==0.17.0
+contourpy==1.3.3
 cryptography==50.0.1
 cuda-bindings==13.3.1
 cuda-core==1.0.1
@@ -27,6 +28,7 @@ cuda-pathfinder==1.8.0
 cuda-python==13.3.1
 cuda-tile==1.5.0
 cuda-toolkit==13.2.1
+cycler==0.12.1
 datasets==5.0.1
 depyf==0.20.0
 detect-installer==0.1.0
@@ -42,6 +44,7 @@ fastar==0.12.0
 fastsafetensors==0.3.3
 filelock==3.32.4
 flashinfer-python==0.6.17
+fonttools==4.64.0
 frozenlist==1.8.0
 fsspec==2026.6.0
 googleapis-common-protos==1.75.2
@@ -57,6 +60,7 @@ huggingface_hub==1.29.0
 humming-kernels==0.1.12
 idna==3.19
 ijson==3.5.1
+iniconfig==2.3.0
 instanttensor==0.1.9
 interegular==0.3.3
 Jinja2==3.1.6
@@ -64,6 +68,7 @@ jiter==0.16.0
 jmespath==1.1.0
 jsonschema==4.26.0
 jsonschema-specifications==2025.9.1
+kiwisolver==1.5.1
 lark==1.2.2
 llguidance==1.7.6
 llvmlite==0.47.0
@@ -71,6 +76,7 @@ lm-format-enforcer==0.11.3
 loguru==0.7.3
 markdown-it-py==4.2.0
 MarkupSafe==3.0.3
+matplotlib==3.11.1
 mcp==2.1.1
 mcp-types==2.1.1
 mdurl==0.1.2
@@ -131,6 +137,7 @@ packaging==26.3
 pandas==3.0.5
 partial-json-parser==0.2.1.1.post7
 pillow==12.3.0
+pluggy==1.6.0
 prometheus-fastapi-instrumentator==8.1.0
 prometheus_client==0.26.0
 propcache==0.5.2
@@ -148,6 +155,8 @@ pydantic_core==2.46.5
 Pygments==2.21.0
 PyJWT==2.13.0
 pynvvideocodec==2.0.4
+pyparsing==3.3.2
+pytest==9.1.1
 python-dateutil==2.9.0.post0
 python-dotenv==1.2.3
 python-json-logger==4.2.0
@@ -175,6 +184,7 @@ starlette==1.6.0
 supervisor==4.3.0
 sympy==1.14.0
 tabulate==0.10.0
+tblib==3.2.2
 tiktoken==0.14.0
 tilelang==0.1.12
 tokenizers==0.23.1
@@ -196,7 +206,7 @@ urllib3==2.7.0
 uv==0.12.7
 uvicorn==0.52.4
 uvloop==0.22.1
--e git+https://github.com/phaelon74/vllm.git@f4cee453bfd38b1afcb8e983517c385e426ec488#egg=vllm
+-e git+https://github.com/phaelon74/vllm.git@c798225a7f197994ed502068fc6ac1ba8a9706f3#egg=vllm
 watchfiles==1.2.0
 websockets==17.1
 xgrammar==0.2.3

environment/repo-git-describe.txt

@@ -1,5 +1,5 @@
 ### git -C /home/phaedawg/kld-bestofbreed/vllm describe --always --dirty --tags
 
-f4cee453b
+c798225a7
 
 ### exit=0

environment/repo-git-log.txt

@@ -1,9 +1,9 @@
 ### git -C /home/phaedawg/kld-bestofbreed/vllm log -n 5 --format=%H %ad %s --date=iso
 
-f4cee453bfd38b1afcb8e983517c385e426ec488 2026-09-02 15:06:34 -0500 Update for Gemma and Qwen QDQ
-b6f11608579c134e3ebfd67c64b34b977d0fc46e 2026-09-02 13:57:15 -0500 qdq fix for NVFP4 attribution
-f623f8a0c8ccdea4f6f9fd69c6bd72aff027b491 2026-09-02 08:20:16 -0500 Update Campaign for Attribute changes.
-4ad4f9caba20804b84e089040e3f0c1f20b8254a 2026-09-02 08:17:15 -0500 Fix Attribution for Qwen in QDQ
-a4b78f8d45751820e035cb37d7e5974a31019bdd 2026-09-02 08:12:58 -0500 Fixed to both Campaign and QDQ as it relates to Attributes for MOE models.
+c798225a7f197994ed502068fc6ac1ba8a9706f3 2026-09-07 20:40:31 -0500 Re-Order Campaign to be more complete.
+fccd81de7ce0e82e7afc23f413cea9fc7908dc5f 2026-09-07 20:16:00 -0500 Treat a report as stale when its bound capture manifest changed
+cbee5c2eb9921e66833f0a3034f314b9915c9250 2026-09-06 13:33:24 -0500 Document W4A4 NVFP4 emulation scoring and probe-based certification
+ae45172d91537370e6d84d6accb0da6a76339bac 2026-09-06 13:23:58 -0500 Pin the W4A4 emulation MoE backend on the student engine only
+575bb21679f76b899018528c9b17159ebf276ab0 2026-09-06 09:40:29 -0500 closed on unprobed MoE backends and score W4A4 NVFP4 faithfully
 
 ### exit=0

environment/runtime.json

@@ -1,16 +1,53 @@
 {
   "capture_runtime_manifest": {
+    "compiled_extensions": {
+      "_C_stable_libtorch.abi3.so": "fbae4a92ec2861cd979eb99a3867049f88f7fcb278337be058dd4cf2ac8c29bf",
+      "_moe_C_stable_libtorch.abi3.so": "a5f41231d3a2fe2958aad06615137ab877671baea20e78b0257516ef3467664e",
+      "third_party/deep_gemm/_C.cpython-310-x86_64-linux-gnu.so": "38b87140577c2d72c2b3a6407c5f07da53ef9362e18fe988162afa8f3f91746e",
+      "third_party/deep_gemm/_C.cpython-311-x86_64-linux-gnu.so": "1f3a3430c188295cf96a5ac977ec72258bccb525daaa7b0ad6642bf3a995d3da",
+      "third_party/deep_gemm/_C.cpython-312-x86_64-linux-gnu.so": "73824dc1312e98cf277ae6bc017becf7e963d5b0bfab81e0c6d475d9b769c9e7",
+      "third_party/deep_gemm/_C.cpython-313-x86_64-linux-gnu.so": "bd0124b39fb500378764bf0d1adb5fe30ea876e66ed65971d8f668de4a8dbe83",
+      "third_party/deep_gemm/_C.cpython-314-x86_64-linux-gnu.so": "0f9817dfe8a619dde56456c6e7a5a9d914fede0fb145bc48aad134e19c77fff0"
+    },
+    "compiled_extensions_sha256": "f2fbc7537b0f01f65341030cfa90741c2ed7ad46c88b0faa3dd0b46d311ccdc8",
     "cuda": "13.2",
+    "determinism": {
+      "CUBLAS_WORKSPACE_CONFIG": ":4096:8",
+      "NCCL_DETERMINISTIC": "1",
+      "VLLM_BATCH_INVARIANT": "1",
+      "VLLM_MOE_USE_DEEP_GEMM": "0"
+    },
     "driver": "580.173.02",
+    "flashinfer": "0.6.17",
     "gpu_names": [
       "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
       "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
       "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
       "NVIDIA RTX PRO 6000 Blackwell Workstation Edition"
     ],
+    "nccl": "2.29.7",
     "platform": "Linux-6.8.0-137-generic-x86_64-with-glibc2.39",
     "python": "3.12.3",
-    "torch": "2.13.0+cu132"
+    "torch": "2.13.0+cu132",
+    "vllm_commit": "c798225a7f197994ed502068fc6ac1ba8a9706f3",
+    "vllm_dirty_digest": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+    "vllm_tree_dirty": false
+  },
+  "compiled_extensions": {
+    "_C_stable_libtorch.abi3.so": "fbae4a92ec2861cd979eb99a3867049f88f7fcb278337be058dd4cf2ac8c29bf",
+    "_moe_C_stable_libtorch.abi3.so": "a5f41231d3a2fe2958aad06615137ab877671baea20e78b0257516ef3467664e",
+    "third_party/deep_gemm/_C.cpython-310-x86_64-linux-gnu.so": "38b87140577c2d72c2b3a6407c5f07da53ef9362e18fe988162afa8f3f91746e",
+    "third_party/deep_gemm/_C.cpython-311-x86_64-linux-gnu.so": "1f3a3430c188295cf96a5ac977ec72258bccb525daaa7b0ad6642bf3a995d3da",
+    "third_party/deep_gemm/_C.cpython-312-x86_64-linux-gnu.so": "73824dc1312e98cf277ae6bc017becf7e963d5b0bfab81e0c6d475d9b769c9e7",
+    "third_party/deep_gemm/_C.cpython-313-x86_64-linux-gnu.so": "bd0124b39fb500378764bf0d1adb5fe30ea876e66ed65971d8f668de4a8dbe83",
+    "third_party/deep_gemm/_C.cpython-314-x86_64-linux-gnu.so": "0f9817dfe8a619dde56456c6e7a5a9d914fede0fb145bc48aad134e19c77fff0"
+  },
+  "compiled_extensions_sha256": "f2fbc7537b0f01f65341030cfa90741c2ed7ad46c88b0faa3dd0b46d311ccdc8",
+  "determinism": {
+    "CUBLAS_WORKSPACE_CONFIG": ":4096:8",
+    "NCCL_DETERMINISTIC": "1",
+    "VLLM_BATCH_INVARIANT": "1",
+    "VLLM_MOE_USE_DEEP_GEMM": "0"
   },
   "devices": [
     {
@@ -43,14 +80,25 @@
     }
   ],
   "env": {
+    "CUBLAS_WORKSPACE_CONFIG": ":4096:8",
     "CUDA_HOME": "/usr/local/cuda-13.0",
-    "HF_TOKEN": "<redacted>"
+    "HF_TOKEN": "<redacted>",
+    "NCCL_DETERMINISTIC": "1",
+    "PYTORCH_NVML_BASED_CUDA_CHECK": "1",
+    "TORCHINDUCTOR_CACHE_DIR": "/tmp/torchinductor_phaedawg",
+    "TORCHINDUCTOR_COMPILE_THREADS": "1",
+    "TRITON_CACHE_AUTOTUNING": "1",
+    "TRITON_PTXAS_BLACKWELL_PATH": "/usr/local/cuda-13.0/bin/ptxas",
+    "VLLM_BATCH_INVARIANT": "1",
+    "VLLM_MARLIN_USE_ATOMIC_ADD": "0",
+    "VLLM_MOE_USE_DEEP_GEMM": "0"
   },
   "env_redacted": [
     "HF_TOKEN"
   ],
   "env_redaction_policy": "Values of variables whose names look like credentials are never recorded. The names are, so a reader can see what was set.",
   "executable": "/home/phaedawg/kld-bestofbreed/kldbestofbreed/bin/python",
+  "flashinfer": "0.6.17",
   "hostname": "d011sd01",
   "platform": "Linux-6.8.0-137-generic-x86_64-with-glibc2.39",
   "python": "3.12.3",
@@ -76,6 +124,7 @@
     "path": "/home/phaedawg/kld-bestofbreed/vllm/vllm/__init__.py",
     "version": "0.1.dev20446+gb2bc9171d"
   },
-  "vllm_commit": "f4cee453bfd38b1afcb8e983517c385e426ec488",
+  "vllm_commit": "c798225a7f197994ed502068fc6ac1ba8a9706f3",
+  "vllm_dirty_digest": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
   "vllm_tree_dirty": false
 }

environment/summary.md

@@ -1,11 +1,11 @@
 # KLD environment report
 
-- captured: 2026-09-02T20:09:33+00:00
+- captured: 2026-09-08T01:43:27+00:00
 - host: d011sd01
 - kernel: 6.8.0-137-generic
 - python: /home/phaedawg/kld-bestofbreed/kldbestofbreed/bin/python
 - repo: /home/phaedawg/kld-bestofbreed/vllm
-- commit: f4cee453bfd38b1afcb8e983517c385e426ec488
+- commit: c798225a7f197994ed502068fc6ac1ba8a9706f3
 - branch: feature/glm53-kld-determinism
 - driver: 580.173.02
 - gpus:       4 NVIDIA RTX PRO 6000 Blackwell Workstation Edition;

excluded-candidates.json

@@ -0,0 +1,11 @@
+{
+  "excluded_candidates": [
+    {
+      "hf_repo": "cklaus/gemma-4-26B-A4B-it-NVFP4",
+      "reason": "Scored on vLLM fdc0b57e while every other candidate scored on c798225a, so its report binds a reference capture the family no longer publishes (Law 12) and it ranks in a comparability group of one. Withdrawn rather than rescored: no measurement of this checkpoint is in question, only the provenance it carries.",
+      "revision": "54218eb0fd33da17632f0d91be841d09a7724a6c"
+    }
+  ],
+  "model": "gemma-4-26B-A4B-it",
+  "schema_version": 1
+}

gemma-4-26B-A4B-it-AWQ-4bit/attribution/gemma-4-26B-A4B-it-qdq-experts-attention-int4_g32_sym-qdq-manifest.json

@@ -227,7 +227,8 @@
     "model": "/media/fmodels2/cyankiwi/gemma-4-26B-A4B-it-AWQ-4bit",
     "quant_algorithm": "round_to_nearest"
   },
-  "note": "Quantize-dequantize variant stored in the source dtype. Weights are numerically what the quantized model would use; arithmetic is not. Comparing this against the real quantized checkpoint isolates the kernel's contribution. This cell matched the int4 format (bit width, group size, symmetry), not a vendor calibration algorithm such as AWQ or GPTQ.",
+  "note": "Quantize-dequantize variant stored in the source dtype. Weights are numerically what the quantized model would use; arithmetic is not. Comparing this against the real quantized checkpoint isolates the kernel's contribution. This cell matched the int4 format (bit width, group size, symmetry), not a vendor calibration algorithm such as AWQ or GPTQ. Reduction dimensions that do not fill the final group were right-zero-padded for rounding and sliced back to source shape.",
+  "padding_policy": "right_zero_pad_reduction_then_slice",
   "quantized_names_sha256": "d0cf67e7b155003d44606ee135bd0dd9b80cccbacafb92001493688dededc673",
   "scheme": "int4_g32_sym",
   "source_model": "/media/fmodels2/google/gemma-4-26B-A4B-it",

gemma-4-26B-A4B-it-AWQ-4bit/attribution/gemma-4-26B-A4B-it-qdq-experts-int4_g32_sym-qdq-manifest.json

@@ -103,7 +103,8 @@
     "model": "/media/fmodels2/cyankiwi/gemma-4-26B-A4B-it-AWQ-4bit",
     "quant_algorithm": "round_to_nearest"
   },
-  "note": "Quantize-dequantize variant stored in the source dtype. Weights are numerically what the quantized model would use; arithmetic is not. Comparing this against the real quantized checkpoint isolates the kernel's contribution. This cell matched the int4 format (bit width, group size, symmetry), not a vendor calibration algorithm such as AWQ or GPTQ.",
+  "note": "Quantize-dequantize variant stored in the source dtype. Weights are numerically what the quantized model would use; arithmetic is not. Comparing this against the real quantized checkpoint isolates the kernel's contribution. This cell matched the int4 format (bit width, group size, symmetry), not a vendor calibration algorithm such as AWQ or GPTQ. Reduction dimensions that do not fill the final group were right-zero-padded for rounding and sliced back to source shape.",
+  "padding_policy": "right_zero_pad_reduction_then_slice",
   "quantized_names_sha256": "fb4cadf81adc3e3200b0e07346c5db64d646173a4526dd15f1b8f6007f43ea16",
   "scheme": "int4_g32_sym",
   "source_model": "/media/fmodels2/google/gemma-4-26B-A4B-it",

gemma-4-26B-A4B-it-AWQ-4bit/attribution/routing/routing_00000.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b15d371f91aff64ce1f741b0fa20fec537834f00753b7e789418d5fbd24ab207
+size 491608

gemma-4-26B-A4B-it-AWQ-4bit/attribution/routing/routing_00001.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:91d03a3baff046afb21017604984267d9abd885408acc1cbcdce2e66784d94c5
+size 491608

gemma-4-26B-A4B-it-AWQ-4bit/attribution/routing/routing_00002.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f60ee2f976b0d66b1f81721901d6a86d51030e7329ef75fa7a6ae11a6f1d98e
+size 491608

gemma-4-26B-A4B-it-AWQ-4bit/attribution/routing/routing_00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45c05ea53d9f540ac7021dc9c914ef086b11b89b15a18e0196a50a2927760f0c
+size 491608

gemma-4-26B-A4B-it-AWQ-4bit/attribution/routing/routing_00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:795504784192f32a321fde2eb7800741d6f841898c0d4ec88cd920cb44730eb0
+size 491608