DmitryDB commited on
Commit
ffe28e2
·
verified ·
1 Parent(s): 91cddc2

Add quality21 Ref2VA checkpoint and validation

Browse files
NOTICE CHANGED
@@ -2,12 +2,12 @@ MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyrig
2
 
3
  Modification notice
4
 
5
- The FL2VA safetensors checkpoint in this repository is a modified derivative of
6
- MiniMax H3. It was converted from the original Diffusers tensor layout to the
7
- ComfyUI layout, selectively quantized to INT8 ConvRot, and compressed with a
8
- 4,097-row FP32 rank-16 table of the AdaLN time-conditioning curve. It uses the
9
- standard MiniMax-H3 table path in ComfyUI and does not require modified runtime
10
- code.
11
 
12
  The two safetensors files under vae/ are unmodified, bit-identical copies of
13
  the Comfy-Org MiniMax-H3 ComfyUI repackages. They retain the MiniMax H3 license
 
2
 
3
  Modification notice
4
 
5
+ The FL2VA and Ref2VA safetensors checkpoints in this repository are modified
6
+ derivatives of MiniMax H3. They were converted from the original Diffusers
7
+ tensor layout to the ComfyUI layout, selectively quantized to INT8 ConvRot, and
8
+ compressed with separate 4,097-row FP32 rank-16 tables of their AdaLN
9
+ time-conditioning curves. They use the standard MiniMax-H3 table path in
10
+ ComfyUI and do not require modified runtime code.
11
 
12
  The two safetensors files under vae/ are unmodified, bit-identical copies of
13
  the Comfy-Org MiniMax-H3 ComfyUI repackages. They retain the MiniMax H3 license
README.md CHANGED
@@ -14,42 +14,46 @@ tags:
14
  - video
15
  - audio
16
  - fl2va
 
17
  ---
18
 
19
  # MiniMax-H3 INT8 Lean ConvRot
20
 
21
  This repository contains a quality-oriented mixed-precision ComfyUI derivative
22
  of [`MiniMaxAI/MiniMax-H3`](https://huggingface.co/MiniMaxAI/MiniMax-H3).
23
- The currently released checkpoint is the **FL2VA** diffusion transformer. It is
24
- designed as a first 24 GiB GPU candidate while preserving more of the original
25
- model than the official pruned INT8 checkpoint.
 
26
 
27
- The release is not an official MiniMax or ComfyOrg checkpoint. It uses the
28
- standard MiniMax-H3 `adaln_t_table` format and loads in an unmodified ComfyUI.
29
 
30
  ## Included files
31
 
32
  | Path | Purpose |
33
  |---|---|
34
  | [`FL2VA/minimax-h3-fl2va-int8-lean-convrot-table-k16-g4097-quality21.safetensors`](FL2VA/minimax-h3-fl2va-int8-lean-convrot-table-k16-g4097-quality21.safetensors) | Stock-compatible FL2VA diffusion transformer, 20.940 GiB |
 
35
  | [`vae/minimax_h3_video_vae_fp16.safetensors`](vae/minimax_h3_video_vae_fp16.safetensors) | Stock ComfyUI MiniMax-H3 visual VAE, FP16, 4.850 GiB |
36
  | [`vae/minimax_h3_audio_vae_fp32.safetensors`](vae/minimax_h3_audio_vae_fp32.safetensors) | Stock ComfyUI MiniMax-H3 stereo audio VAE, FP32, 0.564 GiB |
37
- | [`reports/layer_policy.json`](reports/layer_policy.json) | Exact per-layer precision policy |
38
- | [`reports/validation.json`](reports/validation.json) | Structural, numerical, and CPU-load results |
39
  | [`reports/mm_quant_profile_fl2va.json`](reports/mm_quant_profile_fl2va.json) | Row-sampled reconstruction profile for all 200 main matrices |
 
40
 
41
  This repository intentionally does **not** include the MiniMax-H3 text encoder.
42
  It is being prepared as a separate repository. ComfyUI does not require the
43
  Diffusers scheduler, tokenizer, processor, or remote-code Python bundles as
44
- separate downloads: the standard workflow needs this diffusion model, the two
45
- included VAEs, and a compatible MiniMax-H3 text encoder. Ref2VA is not included
46
- in this release.
47
 
48
  ## What was changed
49
 
50
- The original FL2VA Diffusers shards contain 50 transformer blocks with separate
51
- Q, K, and V tensors. The converter streams those shards directly and performs
52
- the layout changes required by current ComfyUI MiniMax-H3 support:
 
53
 
54
  - packs attention rows as global `cat(Q, K, V)`;
55
  - swaps the two `fc1` halves from `[value, gate]` to `[gate, value]`;
@@ -65,14 +69,14 @@ the layout changes required by current ComfyUI MiniMax-H3 support:
65
  - replaces only the redundant input width of the 51 large AdaLN projections,
66
  while keeping every block and final AdaLN projection separate in FP32.
67
 
68
- The checkpoint contains 872 tensor entries: 170 INT8 weights, 170 quantization
69
- configuration blobs, 250 BF16 entries, and 282 FP32 entries.
70
 
71
  ## Precision policy
72
 
73
  Every DiT block has four main matrices: QKV, attention output, MLP `fc1`, and
74
- MLP `fc2`, for 200 matrices total. The `quality21` profile stores 170 in INT8
75
- and 30 in BF16.
76
 
77
  The BF16 island is:
78
 
@@ -80,11 +84,13 @@ The BF16 island is:
80
  and 49 (27 matrices);
81
  - MLP `fc2` in blocks 39, 45, and 49 (3 matrices).
82
 
83
- These were selected from a deterministic reconstruction audit using 64 output
84
- rows and every input column of each matrix. Attention output projections were
85
- the most error-prone family and are considerably cheaper to retain in BF16 than
86
- QKV or `fc1`. The exact machine-readable list is in
87
- [`reports/layer_policy.json`](reports/layer_policy.json).
 
 
88
 
89
  ## Time conditioning and the AdaLN table
90
 
@@ -108,10 +114,11 @@ Each of the 50 block AdaLN projections and the final AdaLN projection remains a
108
  different FP32 linear layer. Only their shared input curve is compressed.
109
 
110
  On 19 validation timesteps, the relative AdaLN output error against the raw HF
111
- weights was `3.04e-7` to `3.82e-7` for blocks 0, 24, 49, and the final layer.
112
- For context, the official pruned rank-8 FP16 table measured about
113
- `0.016%–0.021%` on the same projections. This comparison covers only the
114
- time/AdaLN path, not end-to-end generation quality.
 
115
 
116
  ## Comparison with official ComfyOrg files
117
 
@@ -119,38 +126,44 @@ time/AdaLN path, not end-to-end generation quality.
119
  |---|---:|---:|---|
120
  | Official regular INT8 ConvRot | about 31.70 GiB | 200 INT8 | full AdaLN weights |
121
  | Official pruned INT8 ConvRot | 19.530 GiB | 200 INT8 | FP16 rank-8 table |
122
- | This `quality21` release | 20.940 GiB | 170 INT8 + 30 BF16 | FP32 rank-16 table, 4,097 rows |
123
 
124
- This file is 1.410 GiB (7.22%) larger than the official pruned INT8 checkpoint.
125
- The added space is intentionally spent on a much denser, higher-precision time
126
- curve and a BF16 island for the most sensitive main matrices.
 
127
 
128
- The full 200-matrix sampled reconstruction proxy produced:
129
 
130
- | Metric | Official-style absmax INT8 | This profile |
131
  |---|---:|---:|
132
- | Unweighted mean relative matrix error | 0.9345% | 0.7639% |
 
133
 
134
- For this profile, selected BF16 matrices count as zero storage reconstruction
135
  error. On the remaining 170 INT8 matrices, searched scales reduced the mean
136
- proxy from 0.9109% to 0.8987%. The overall 18.26% reduction is a sampled weight
137
- reconstruction result, **not** a perceptual or generation benchmark.
 
138
 
139
  ## ComfyUI installation
140
 
141
  1. Use a current ComfyUI checkout with MiniMax-H3 and native INT8 ConvRot
142
  support. This release was CPU-load-tested on a clean, unmodified ComfyUI
143
  commit `14b05228cef127ce529bc0c08660770d4af3e9a8`.
144
- 2. Place the FL2VA safetensors file in
145
- `ComfyUI/models/diffusion_models/`.
 
146
  3. Place both files from `vae/` in `ComfyUI/models/vae/`.
147
- 4. Load the three files with the standard diffusion-model and VAE loaders in a
148
- MiniMax-H3 FL2VA workflow. Supply a compatible text encoder separately.
 
149
 
150
  ```text
151
  ComfyUI/models/
152
  |-- diffusion_models/
153
- | `-- minimax-h3-fl2va-int8-lean-convrot-table-k16-g4097-quality21.safetensors
 
154
  |-- vae/
155
  | |-- minimax_h3_video_vae_fp16.safetensors
156
  | `-- minimax_h3_audio_vae_fp32.safetensors
@@ -158,21 +171,28 @@ ComfyUI/models/
158
  `-- <MiniMax-H3 text encoder from its separate repository>
159
  ```
160
 
161
- No custom node or core patch is required for this checkpoint.
162
 
163
  ## Validation status
164
 
165
  Completed:
166
 
167
  - safetensors structure, shapes, dtypes, and all 170 ConvRot descriptors;
168
- - SHA-256: `ad0cf8b50ba92068b5a4477527d5da73412a22db99cded35b07195504e78cf6a`;
 
 
 
169
  - BF16 main island: 30/30 matrices bitwise equal to raw HF;
170
- - 871/871 tensors outside the time representation preserved bit-for-bit during
171
- the final stock-table conversion;
172
- - FP32 table grid reproduced exactly and interpolation checked at every one of
173
- the 4,096 interval midpoints;
 
 
 
174
  - full CPU load through a clean, unmodified ComfyUI as
175
- `ModelPatcher -> MiniMaxH3Model`, table mode `(4097, 16)`.
 
176
  - both companion VAE files fully loaded on CPU through the same clean ComfyUI:
177
  `MiniMaxH3VideoVAE` in FP16 with native internal tiling, and
178
  `MiniMaxH3AudioVAE` in FP32 with 32 kHz stereo output;
@@ -184,9 +204,9 @@ Completed:
184
 
185
  Not completed at release time:
186
 
187
- - CUDA/GPU loading of this exact checkpoint;
188
  - peak VRAM measurement for a target resolution and frame count;
189
- - end-to-end FL2VA generation;
190
  - same-seed visual/audio A/B against BF16 and official pruned variants.
191
 
192
  Therefore, 20.940 GiB is the on-disk tensor payload, not a promise that every
@@ -195,15 +215,18 @@ frame count, batch size, and offloading policy determine peak VRAM.
195
 
196
  ## Conversion provenance
197
 
198
- The converter reads the original top-level `transformer/` Diffusers shards from
199
- MiniMaxAI directly; it does not create a 61+ GiB merged BF16 intermediate. The
200
- nested task-specific transformer folder was not used because its already-packed
201
- QKV is head-major, while current ComfyUI expects global `cat(Q,K,V)` packing.
 
202
 
203
- The official
204
  [`minimax_h3_fl2va_pruned_int8_convrot.safetensors`](https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors)
205
- was used only as a layout reference and as the source of the FP32
206
- `rope.inv_freq` tensor absent from the raw Diffusers state dict.
 
 
207
 
208
  The two files under `vae/` are unmodified, bit-identical copies of the standard
209
  [`Comfy-Org/MiniMax-H3`](https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/vae)
@@ -220,8 +243,9 @@ latent statistics. No remote Python code or external VAE config is required.
220
 
221
  This release is intended for research and local ComfyUI inference. It is a
222
  new, structurally validated operating point, not a completed quality benchmark.
223
- The current layer policy was profiled on FL2VA and must not be assumed optimal
224
- for Ref2VA without a separate audit.
 
225
 
226
  ## License
227
 
 
14
  - video
15
  - audio
16
  - fl2va
17
+ - ref2va
18
  ---
19
 
20
  # MiniMax-H3 INT8 Lean ConvRot
21
 
22
  This repository contains a quality-oriented mixed-precision ComfyUI derivative
23
  of [`MiniMaxAI/MiniMax-H3`](https://huggingface.co/MiniMaxAI/MiniMax-H3).
24
+ The released checkpoints cover both the **FL2VA** and **Ref2VA** diffusion
25
+ transformers. Each is designed as a first 24 GiB GPU candidate while preserving
26
+ more of the original model than the corresponding official pruned INT8
27
+ checkpoint.
28
 
29
+ These are not official MiniMax or ComfyOrg checkpoints. They use the standard
30
+ MiniMax-H3 `adaln_t_table` format and load in an unmodified ComfyUI.
31
 
32
  ## Included files
33
 
34
  | Path | Purpose |
35
  |---|---|
36
  | [`FL2VA/minimax-h3-fl2va-int8-lean-convrot-table-k16-g4097-quality21.safetensors`](FL2VA/minimax-h3-fl2va-int8-lean-convrot-table-k16-g4097-quality21.safetensors) | Stock-compatible FL2VA diffusion transformer, 20.940 GiB |
37
+ | [`Ref2VA/minimax-h3-ref2va-int8-lean-convrot-table-k16-g4097-quality21.safetensors`](Ref2VA/minimax-h3-ref2va-int8-lean-convrot-table-k16-g4097-quality21.safetensors) | Stock-compatible Ref2VA diffusion transformer, 20.940 GiB |
38
  | [`vae/minimax_h3_video_vae_fp16.safetensors`](vae/minimax_h3_video_vae_fp16.safetensors) | Stock ComfyUI MiniMax-H3 visual VAE, FP16, 4.850 GiB |
39
  | [`vae/minimax_h3_audio_vae_fp32.safetensors`](vae/minimax_h3_audio_vae_fp32.safetensors) | Stock ComfyUI MiniMax-H3 stereo audio VAE, FP32, 0.564 GiB |
40
+ | [`reports/layer_policy.json`](reports/layer_policy.json) / [`layer_policy_ref2va.json`](reports/layer_policy_ref2va.json) | Exact per-layer precision policies |
41
+ | [`reports/validation.json`](reports/validation.json) / [`validation_ref2va.json`](reports/validation_ref2va.json) | Structural, numerical, and CPU-load results |
42
  | [`reports/mm_quant_profile_fl2va.json`](reports/mm_quant_profile_fl2va.json) | Row-sampled reconstruction profile for all 200 main matrices |
43
+ | [`reports/mm_quant_profile_ref2va.json`](reports/mm_quant_profile_ref2va.json) | Matching Ref2VA reconstruction profile |
44
 
45
  This repository intentionally does **not** include the MiniMax-H3 text encoder.
46
  It is being prepared as a separate repository. ComfyUI does not require the
47
  Diffusers scheduler, tokenizer, processor, or remote-code Python bundles as
48
+ separate downloads: a standard workflow needs one selected diffusion model,
49
+ the two included VAEs, and a compatible MiniMax-H3 text encoder.
 
50
 
51
  ## What was changed
52
 
53
+ The original FL2VA and Ref2VA Diffusers shards each contain 50 transformer
54
+ blocks with separate Q, K, and V tensors. The converter streams the selected
55
+ variant directly and performs the layout changes required by current ComfyUI
56
+ MiniMax-H3 support:
57
 
58
  - packs attention rows as global `cat(Q, K, V)`;
59
  - swaps the two `fc1` halves from `[value, gate]` to `[gate, value]`;
 
69
  - replaces only the redundant input width of the 51 large AdaLN projections,
70
  while keeping every block and final AdaLN projection separate in FP32.
71
 
72
+ Each checkpoint contains 872 tensor entries: 170 INT8 weights, 170
73
+ quantization configuration blobs, 250 BF16 entries, and 282 FP32 entries.
74
 
75
  ## Precision policy
76
 
77
  Every DiT block has four main matrices: QKV, attention output, MLP `fc1`, and
78
+ MLP `fc2`, for 200 matrices total per checkpoint. The `quality21` profile stores
79
+ 170 in INT8 and 30 in BF16.
80
 
81
  The BF16 island is:
82
 
 
84
  and 49 (27 matrices);
85
  - MLP `fc2` in blocks 39, 45, and 49 (3 matrices).
86
 
87
+ These were selected from deterministic reconstruction audits using 64 output
88
+ rows and every input column of each matrix. Ref2VA independently selected the
89
+ same 30-layer set as FL2VA, with a per-layer error correlation of `0.999676`.
90
+ Attention output projections were the most error-prone family and are
91
+ considerably cheaper to retain in BF16 than QKV or `fc1`. The exact
92
+ machine-readable lists are in [`reports/layer_policy.json`](reports/layer_policy.json)
93
+ and [`reports/layer_policy_ref2va.json`](reports/layer_policy_ref2va.json).
94
 
95
  ## Time conditioning and the AdaLN table
96
 
 
114
  different FP32 linear layer. Only their shared input curve is compressed.
115
 
116
  On 19 validation timesteps, the relative AdaLN output error against the raw HF
117
+ weights was `3.04e-7` to `3.82e-7` for FL2VA and `3.42e-7` to `4.46e-7` for
118
+ Ref2VA across blocks 0, 24, 49, and the final layer. For context, the official
119
+ pruned rank-8 FP16 tables measured about `0.016%–0.021%` on the same
120
+ projections. This comparison covers only the time/AdaLN path, not end-to-end
121
+ generation quality.
122
 
123
  ## Comparison with official ComfyOrg files
124
 
 
126
  |---|---:|---:|---|
127
  | Official regular INT8 ConvRot | about 31.70 GiB | 200 INT8 | full AdaLN weights |
128
  | Official pruned INT8 ConvRot | 19.530 GiB | 200 INT8 | FP16 rank-8 table |
129
+ | Each `quality21` release | 20.940 GiB | 170 INT8 + 30 BF16 | FP32 rank-16 table, 4,097 rows |
130
 
131
+ Each file is about 1.410 GiB (7.22%) larger than its official pruned INT8
132
+ counterpart. The added space is intentionally spent on a much denser,
133
+ higher-precision time curve and a BF16 island for the most sensitive main
134
+ matrices.
135
 
136
+ The full 200-matrix sampled reconstruction proxies produced:
137
 
138
+ | Variant | Official-style absmax INT8 | This profile |
139
  |---|---:|---:|
140
+ | FL2VA | 0.9345% | 0.7639% |
141
+ | Ref2VA | 0.9345% | 0.7639% |
142
 
143
+ For these profiles, selected BF16 matrices count as zero storage reconstruction
144
  error. On the remaining 170 INT8 matrices, searched scales reduced the mean
145
+ proxy from about 0.9109% to 0.8987%. The approximately 18.25% overall reduction
146
+ is a sampled weight reconstruction result, **not** a perceptual or generation
147
+ benchmark.
148
 
149
  ## ComfyUI installation
150
 
151
  1. Use a current ComfyUI checkout with MiniMax-H3 and native INT8 ConvRot
152
  support. This release was CPU-load-tested on a clean, unmodified ComfyUI
153
  commit `14b05228cef127ce529bc0c08660770d4af3e9a8`.
154
+ 2. Place the FL2VA and/or Ref2VA safetensors file in
155
+ `ComfyUI/models/diffusion_models/`. FL2VA is for text/image first-last-frame
156
+ workflows; Ref2VA is for reference image/video/audio workflows.
157
  3. Place both files from `vae/` in `ComfyUI/models/vae/`.
158
+ 4. Load one diffusion checkpoint and both VAEs with the standard loaders in the
159
+ matching MiniMax-H3 FL2VA or Ref2VA workflow. Supply a compatible text
160
+ encoder separately.
161
 
162
  ```text
163
  ComfyUI/models/
164
  |-- diffusion_models/
165
+ | |-- minimax-h3-fl2va-int8-lean-convrot-table-k16-g4097-quality21.safetensors
166
+ | `-- minimax-h3-ref2va-int8-lean-convrot-table-k16-g4097-quality21.safetensors
167
  |-- vae/
168
  | |-- minimax_h3_video_vae_fp16.safetensors
169
  | `-- minimax_h3_audio_vae_fp32.safetensors
 
171
  `-- <MiniMax-H3 text encoder from its separate repository>
172
  ```
173
 
174
+ No custom node or core patch is required for either checkpoint.
175
 
176
  ## Validation status
177
 
178
  Completed:
179
 
180
  - safetensors structure, shapes, dtypes, and all 170 ConvRot descriptors;
181
+ - SHA-256: FL2VA
182
+ `ad0cf8b50ba92068b5a4477527d5da73412a22db99cded35b07195504e78cf6a`;
183
+ Ref2VA
184
+ `f4b2a8b4a53446f5d696a39deb693b09ac7ce37013793a8b19340644e38b154f`;
185
  - BF16 main island: 30/30 matrices bitwise equal to raw HF;
186
+ - FL2VA: 871/871 tensors outside the time representation preserved bit-for-bit
187
+ during the final stock-table conversion;
188
+ - Ref2VA: 258/258 direct source-copy tensors (4,146,503,680 bytes) bitwise equal
189
+ to raw HF;
190
+ - FP32 table interpolation checked at every one of the 4,096 interval
191
+ midpoints; FL2VA grid rows reproduced exactly, and the independent Ref2VA
192
+ FP64-SVD rebuild differed by at most `4.55e-13`;
193
  - full CPU load through a clean, unmodified ComfyUI as
194
+ `ModelPatcher -> MiniMaxH3Model`, table mode `(4097, 16)`, for both diffusion
195
+ checkpoints;
196
  - both companion VAE files fully loaded on CPU through the same clean ComfyUI:
197
  `MiniMaxH3VideoVAE` in FP16 with native internal tiling, and
198
  `MiniMaxH3AudioVAE` in FP32 with 32 kHz stereo output;
 
204
 
205
  Not completed at release time:
206
 
207
+ - CUDA/GPU loading of these exact checkpoints;
208
  - peak VRAM measurement for a target resolution and frame count;
209
+ - end-to-end FL2VA or Ref2VA generation;
210
  - same-seed visual/audio A/B against BF16 and official pruned variants.
211
 
212
  Therefore, 20.940 GiB is the on-disk tensor payload, not a promise that every
 
215
 
216
  ## Conversion provenance
217
 
218
+ The converter reads the original top-level `transformer/` or `transformer_ref/`
219
+ Diffusers shards from MiniMaxAI directly; it does not create a 61+ GiB merged
220
+ BF16 intermediate. The nested task-specific transformer folders were not used
221
+ because their already-packed QKV is head-major, while current ComfyUI expects
222
+ global `cat(Q,K,V)` packing.
223
 
224
+ The official FL2VA and Ref2VA pruned files,
225
  [`minimax_h3_fl2va_pruned_int8_convrot.safetensors`](https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors)
226
+ and
227
+ [`minimax_h3_ref2va_pruned_int8_convrot.safetensors`](https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors),
228
+ were used only as layout references and as the sources of the corresponding
229
+ FP32 `rope.inv_freq` tensors absent from the raw Diffusers state dicts.
230
 
231
  The two files under `vae/` are unmodified, bit-identical copies of the standard
232
  [`Comfy-Org/MiniMax-H3`](https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/vae)
 
243
 
244
  This release is intended for research and local ComfyUI inference. It is a
245
  new, structurally validated operating point, not a completed quality benchmark.
246
+ Both variants were profiled separately. Their independently selected
247
+ quality21 BF16 islands are identical, but this does not establish perceptual
248
+ quality without end-to-end A/B generation.
249
 
250
  ## License
251
 
reports/layer_policy_ref2va.json ADDED
@@ -0,0 +1,57 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "variant": "Ref2VA",
3
+ "profile": "quality21",
4
+ "main_matrix_count": 200,
5
+ "int8_convrot_group_256_count": 170,
6
+ "bf16_main_count": 30,
7
+ "bf16_main_layers": [
8
+ "blocks.0.attn.out_proj",
9
+ "blocks.1.attn.out_proj",
10
+ "blocks.2.attn.out_proj",
11
+ "blocks.3.attn.out_proj",
12
+ "blocks.4.attn.out_proj",
13
+ "blocks.5.attn.out_proj",
14
+ "blocks.6.attn.out_proj",
15
+ "blocks.7.attn.out_proj",
16
+ "blocks.8.attn.out_proj",
17
+ "blocks.9.attn.out_proj",
18
+ "blocks.10.attn.out_proj",
19
+ "blocks.11.attn.out_proj",
20
+ "blocks.12.attn.out_proj",
21
+ "blocks.13.attn.out_proj",
22
+ "blocks.14.attn.out_proj",
23
+ "blocks.15.attn.out_proj",
24
+ "blocks.17.attn.out_proj",
25
+ "blocks.19.attn.out_proj",
26
+ "blocks.20.attn.out_proj",
27
+ "blocks.27.attn.out_proj",
28
+ "blocks.38.attn.out_proj",
29
+ "blocks.43.attn.out_proj",
30
+ "blocks.44.attn.out_proj",
31
+ "blocks.45.attn.out_proj",
32
+ "blocks.46.attn.out_proj",
33
+ "blocks.47.attn.out_proj",
34
+ "blocks.49.attn.out_proj",
35
+ "blocks.39.mlp.fc2",
36
+ "blocks.45.mlp.fc2",
37
+ "blocks.49.mlp.fc2"
38
+ ],
39
+ "selection_validation": {
40
+ "method": "deterministic 64 output rows per matrix, all input columns",
41
+ "fl2va_ref2va_error_correlation": 0.9996760487556458,
42
+ "same_layers_as_fl2va": true
43
+ },
44
+ "other_precision": {
45
+ "token_refiner": "source BF16",
46
+ "normalization_and_qk_norm": "source BF16",
47
+ "condition_projection": "source BF16",
48
+ "video_audio_patch_and_output_projections": "source FP32",
49
+ "adaln_projections": "51 separate FP32 rank-16 projections",
50
+ "adaln_time_table": "FP32 [4097, 16], generated from the original Ref2VA FP32 time embedder and shared rank-16 basis",
51
+ "rope_inv_freq": "FP32 [16]"
52
+ },
53
+ "packing": {
54
+ "qkv": "global cat(Q,K,V) for current ComfyUI MiniMax-H3 split",
55
+ "mlp_fc1": "halves reordered from [value,gate] to [gate,value]"
56
+ }
57
+ }
reports/mm_quant_profile_ref2va.json ADDED
@@ -0,0 +1,3011 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "variant": "ref2va",
3
+ "source": "F:\\AI\\hf_originals\\MiniMax__MiniMax-H3\\HF\\transformer_ref",
4
+ "groupsize": 256,
5
+ "rows": 64,
6
+ "search_min": 0.8,
7
+ "clip_margin": 0.05,
8
+ "elapsed_seconds": 66.3423388004303,
9
+ "layers": [
10
+ {
11
+ "layer": "blocks.0.attn.qkv_proj",
12
+ "block": 0,
13
+ "kind": "attn.qkv_proj",
14
+ "shape": [
15
+ 21504,
16
+ 5376
17
+ ],
18
+ "sample_rows": 64,
19
+ "absmax_error": 0.009149877354502678,
20
+ "searched_error": 0.009098502807319164,
21
+ "improvement": 5.137454718351364e-05,
22
+ "clipped_row_fraction": 0.109375,
23
+ "bf16_delta_bytes": 115519416
24
+ },
25
+ {
26
+ "layer": "blocks.0.attn.out_proj",
27
+ "block": 0,
28
+ "kind": "attn.out_proj",
29
+ "shape": [
30
+ 5376,
31
+ 7168
32
+ ],
33
+ "sample_rows": 64,
34
+ "absmax_error": 0.011861145496368408,
35
+ "searched_error": 0.01176026277244091,
36
+ "improvement": 0.00010088272392749786,
37
+ "clipped_row_fraction": 0.3125,
38
+ "bf16_delta_bytes": 38513592
39
+ },
40
+ {
41
+ "layer": "blocks.0.mlp.fc1",
42
+ "block": 0,
43
+ "kind": "mlp.fc1",
44
+ "shape": [
45
+ 28672,
46
+ 5376
47
+ ],
48
+ "sample_rows": 64,
49
+ "absmax_error": 0.008755319751799107,
50
+ "searched_error": 0.00868719257414341,
51
+ "improvement": 6.812717765569687e-05,
52
+ "clipped_row_fraction": 0.21875,
53
+ "bf16_delta_bytes": 154025912
54
+ },
55
+ {
56
+ "layer": "blocks.0.mlp.fc2",
57
+ "block": 0,
58
+ "kind": "mlp.fc2",
59
+ "shape": [
60
+ 5376,
61
+ 14336
62
+ ],
63
+ "sample_rows": 64,
64
+ "absmax_error": 0.009267843328416348,
65
+ "searched_error": 0.00897966418415308,
66
+ "improvement": 0.0002881791442632675,
67
+ "clipped_row_fraction": 0.859375,
68
+ "bf16_delta_bytes": 77048760
69
+ },
70
+ {
71
+ "layer": "blocks.1.attn.qkv_proj",
72
+ "block": 1,
73
+ "kind": "attn.qkv_proj",
74
+ "shape": [
75
+ 21504,
76
+ 5376
77
+ ],
78
+ "sample_rows": 64,
79
+ "absmax_error": 0.008966601453721523,
80
+ "searched_error": 0.008890115655958652,
81
+ "improvement": 7.648579776287079e-05,
82
+ "clipped_row_fraction": 0.171875,
83
+ "bf16_delta_bytes": 115519416
84
+ },
85
+ {
86
+ "layer": "blocks.1.attn.out_proj",
87
+ "block": 1,
88
+ "kind": "attn.out_proj",
89
+ "shape": [
90
+ 5376,
91
+ 7168
92
+ ],
93
+ "sample_rows": 64,
94
+ "absmax_error": 0.012773617170751095,
95
+ "searched_error": 0.012632046826183796,
96
+ "improvement": 0.0001415703445672989,
97
+ "clipped_row_fraction": 0.390625,
98
+ "bf16_delta_bytes": 38513592
99
+ },
100
+ {
101
+ "layer": "blocks.1.mlp.fc1",
102
+ "block": 1,
103
+ "kind": "mlp.fc1",
104
+ "shape": [
105
+ 28672,
106
+ 5376
107
+ ],
108
+ "sample_rows": 64,
109
+ "absmax_error": 0.00876633170992136,
110
+ "searched_error": 0.008713556453585625,
111
+ "improvement": 5.277525633573532e-05,
112
+ "clipped_row_fraction": 0.171875,
113
+ "bf16_delta_bytes": 154025912
114
+ },
115
+ {
116
+ "layer": "blocks.1.mlp.fc2",
117
+ "block": 1,
118
+ "kind": "mlp.fc2",
119
+ "shape": [
120
+ 5376,
121
+ 14336
122
+ ],
123
+ "sample_rows": 64,
124
+ "absmax_error": 0.009261968545615673,
125
+ "searched_error": 0.008988389745354652,
126
+ "improvement": 0.00027357880026102066,
127
+ "clipped_row_fraction": 0.828125,
128
+ "bf16_delta_bytes": 77048760
129
+ },
130
+ {
131
+ "layer": "blocks.2.attn.qkv_proj",
132
+ "block": 2,
133
+ "kind": "attn.qkv_proj",
134
+ "shape": [
135
+ 21504,
136
+ 5376
137
+ ],
138
+ "sample_rows": 64,
139
+ "absmax_error": 0.008780846372246742,
140
+ "searched_error": 0.00874966662377119,
141
+ "improvement": 3.1179748475551605e-05,
142
+ "clipped_row_fraction": 0.109375,
143
+ "bf16_delta_bytes": 115519416
144
+ },
145
+ {
146
+ "layer": "blocks.2.attn.out_proj",
147
+ "block": 2,
148
+ "kind": "attn.out_proj",
149
+ "shape": [
150
+ 5376,
151
+ 7168
152
+ ],
153
+ "sample_rows": 64,
154
+ "absmax_error": 0.012882267124950886,
155
+ "searched_error": 0.01274180319160223,
156
+ "improvement": 0.0001404639333486557,
157
+ "clipped_row_fraction": 0.328125,
158
+ "bf16_delta_bytes": 38513592
159
+ },
160
+ {
161
+ "layer": "blocks.2.mlp.fc1",
162
+ "block": 2,
163
+ "kind": "mlp.fc1",
164
+ "shape": [
165
+ 28672,
166
+ 5376
167
+ ],
168
+ "sample_rows": 64,
169
+ "absmax_error": 0.008909977041184902,
170
+ "searched_error": 0.008883747272193432,
171
+ "improvement": 2.6229768991470337e-05,
172
+ "clipped_row_fraction": 0.109375,
173
+ "bf16_delta_bytes": 154025912
174
+ },
175
+ {
176
+ "layer": "blocks.2.mlp.fc2",
177
+ "block": 2,
178
+ "kind": "mlp.fc2",
179
+ "shape": [
180
+ 5376,
181
+ 14336
182
+ ],
183
+ "sample_rows": 64,
184
+ "absmax_error": 0.009286819025874138,
185
+ "searched_error": 0.009042869322001934,
186
+ "improvement": 0.00024394970387220383,
187
+ "clipped_row_fraction": 0.765625,
188
+ "bf16_delta_bytes": 77048760
189
+ },
190
+ {
191
+ "layer": "blocks.3.attn.qkv_proj",
192
+ "block": 3,
193
+ "kind": "attn.qkv_proj",
194
+ "shape": [
195
+ 21504,
196
+ 5376
197
+ ],
198
+ "sample_rows": 64,
199
+ "absmax_error": 0.008728427812457085,
200
+ "searched_error": 0.008676868863403797,
201
+ "improvement": 5.1558949053287506e-05,
202
+ "clipped_row_fraction": 0.21875,
203
+ "bf16_delta_bytes": 115519416
204
+ },
205
+ {
206
+ "layer": "blocks.3.attn.out_proj",
207
+ "block": 3,
208
+ "kind": "attn.out_proj",
209
+ "shape": [
210
+ 5376,
211
+ 7168
212
+ ],
213
+ "sample_rows": 64,
214
+ "absmax_error": 0.011711601167917252,
215
+ "searched_error": 0.011610935442149639,
216
+ "improvement": 0.00010066572576761246,
217
+ "clipped_row_fraction": 0.359375,
218
+ "bf16_delta_bytes": 38513592
219
+ },
220
+ {
221
+ "layer": "blocks.3.mlp.fc1",
222
+ "block": 3,
223
+ "kind": "mlp.fc1",
224
+ "shape": [
225
+ 28672,
226
+ 5376
227
+ ],
228
+ "sample_rows": 64,
229
+ "absmax_error": 0.008689207956194878,
230
+ "searched_error": 0.008630858734250069,
231
+ "improvement": 5.834922194480896e-05,
232
+ "clipped_row_fraction": 0.234375,
233
+ "bf16_delta_bytes": 154025912
234
+ },
235
+ {
236
+ "layer": "blocks.3.mlp.fc2",
237
+ "block": 3,
238
+ "kind": "mlp.fc2",
239
+ "shape": [
240
+ 5376,
241
+ 14336
242
+ ],
243
+ "sample_rows": 64,
244
+ "absmax_error": 0.009451334364712238,
245
+ "searched_error": 0.00917091965675354,
246
+ "improvement": 0.00028041470795869827,
247
+ "clipped_row_fraction": 0.796875,
248
+ "bf16_delta_bytes": 77048760
249
+ },
250
+ {
251
+ "layer": "blocks.4.attn.qkv_proj",
252
+ "block": 4,
253
+ "kind": "attn.qkv_proj",
254
+ "shape": [
255
+ 21504,
256
+ 5376
257
+ ],
258
+ "sample_rows": 64,
259
+ "absmax_error": 0.008999093435704708,
260
+ "searched_error": 0.008943070657551289,
261
+ "improvement": 5.6022778153419495e-05,
262
+ "clipped_row_fraction": 0.171875,
263
+ "bf16_delta_bytes": 115519416
264
+ },
265
+ {
266
+ "layer": "blocks.4.attn.out_proj",
267
+ "block": 4,
268
+ "kind": "attn.out_proj",
269
+ "shape": [
270
+ 5376,
271
+ 7168
272
+ ],
273
+ "sample_rows": 64,
274
+ "absmax_error": 0.010633870959281921,
275
+ "searched_error": 0.01052608247846365,
276
+ "improvement": 0.00010778848081827164,
277
+ "clipped_row_fraction": 0.265625,
278
+ "bf16_delta_bytes": 38513592
279
+ },
280
+ {
281
+ "layer": "blocks.4.mlp.fc1",
282
+ "block": 4,
283
+ "kind": "mlp.fc1",
284
+ "shape": [
285
+ 28672,
286
+ 5376
287
+ ],
288
+ "sample_rows": 64,
289
+ "absmax_error": 0.008890100754797459,
290
+ "searched_error": 0.00882396474480629,
291
+ "improvement": 6.613600999116898e-05,
292
+ "clipped_row_fraction": 0.265625,
293
+ "bf16_delta_bytes": 154025912
294
+ },
295
+ {
296
+ "layer": "blocks.4.mlp.fc2",
297
+ "block": 4,
298
+ "kind": "mlp.fc2",
299
+ "shape": [
300
+ 5376,
301
+ 14336
302
+ ],
303
+ "sample_rows": 64,
304
+ "absmax_error": 0.009374679997563362,
305
+ "searched_error": 0.00907217524945736,
306
+ "improvement": 0.0003025047481060028,
307
+ "clipped_row_fraction": 0.875,
308
+ "bf16_delta_bytes": 77048760
309
+ },
310
+ {
311
+ "layer": "blocks.5.attn.qkv_proj",
312
+ "block": 5,
313
+ "kind": "attn.qkv_proj",
314
+ "shape": [
315
+ 21504,
316
+ 5376
317
+ ],
318
+ "sample_rows": 64,
319
+ "absmax_error": 0.008881070651113987,
320
+ "searched_error": 0.00881462637335062,
321
+ "improvement": 6.64442777633667e-05,
322
+ "clipped_row_fraction": 0.265625,
323
+ "bf16_delta_bytes": 115519416
324
+ },
325
+ {
326
+ "layer": "blocks.5.attn.out_proj",
327
+ "block": 5,
328
+ "kind": "attn.out_proj",
329
+ "shape": [
330
+ 5376,
331
+ 7168
332
+ ],
333
+ "sample_rows": 64,
334
+ "absmax_error": 0.011126119643449783,
335
+ "searched_error": 0.011061684228479862,
336
+ "improvement": 6.443541496992111e-05,
337
+ "clipped_row_fraction": 0.234375,
338
+ "bf16_delta_bytes": 38513592
339
+ },
340
+ {
341
+ "layer": "blocks.5.mlp.fc1",
342
+ "block": 5,
343
+ "kind": "mlp.fc1",
344
+ "shape": [
345
+ 28672,
346
+ 5376
347
+ ],
348
+ "sample_rows": 64,
349
+ "absmax_error": 0.008905366994440556,
350
+ "searched_error": 0.00883188284933567,
351
+ "improvement": 7.34841451048851e-05,
352
+ "clipped_row_fraction": 0.265625,
353
+ "bf16_delta_bytes": 154025912
354
+ },
355
+ {
356
+ "layer": "blocks.5.mlp.fc2",
357
+ "block": 5,
358
+ "kind": "mlp.fc2",
359
+ "shape": [
360
+ 5376,
361
+ 14336
362
+ ],
363
+ "sample_rows": 64,
364
+ "absmax_error": 0.009375902824103832,
365
+ "searched_error": 0.009108574129641056,
366
+ "improvement": 0.0002673286944627762,
367
+ "clipped_row_fraction": 0.75,
368
+ "bf16_delta_bytes": 77048760
369
+ },
370
+ {
371
+ "layer": "blocks.6.attn.qkv_proj",
372
+ "block": 6,
373
+ "kind": "attn.qkv_proj",
374
+ "shape": [
375
+ 21504,
376
+ 5376
377
+ ],
378
+ "sample_rows": 64,
379
+ "absmax_error": 0.008814658969640732,
380
+ "searched_error": 0.008784820325672626,
381
+ "improvement": 2.9838643968105316e-05,
382
+ "clipped_row_fraction": 0.125,
383
+ "bf16_delta_bytes": 115519416
384
+ },
385
+ {
386
+ "layer": "blocks.6.attn.out_proj",
387
+ "block": 6,
388
+ "kind": "attn.out_proj",
389
+ "shape": [
390
+ 5376,
391
+ 7168
392
+ ],
393
+ "sample_rows": 64,
394
+ "absmax_error": 0.010695223696529865,
395
+ "searched_error": 0.010601626709103584,
396
+ "improvement": 9.359698742628098e-05,
397
+ "clipped_row_fraction": 0.3125,
398
+ "bf16_delta_bytes": 38513592
399
+ },
400
+ {
401
+ "layer": "blocks.6.mlp.fc1",
402
+ "block": 6,
403
+ "kind": "mlp.fc1",
404
+ "shape": [
405
+ 28672,
406
+ 5376
407
+ ],
408
+ "sample_rows": 64,
409
+ "absmax_error": 0.008920718915760517,
410
+ "searched_error": 0.008867485448718071,
411
+ "improvement": 5.3233467042446136e-05,
412
+ "clipped_row_fraction": 0.25,
413
+ "bf16_delta_bytes": 154025912
414
+ },
415
+ {
416
+ "layer": "blocks.6.mlp.fc2",
417
+ "block": 6,
418
+ "kind": "mlp.fc2",
419
+ "shape": [
420
+ 5376,
421
+ 14336
422
+ ],
423
+ "sample_rows": 64,
424
+ "absmax_error": 0.009407012723386288,
425
+ "searched_error": 0.00913174543529749,
426
+ "improvement": 0.0002752672880887985,
427
+ "clipped_row_fraction": 0.8125,
428
+ "bf16_delta_bytes": 77048760
429
+ },
430
+ {
431
+ "layer": "blocks.7.attn.qkv_proj",
432
+ "block": 7,
433
+ "kind": "attn.qkv_proj",
434
+ "shape": [
435
+ 21504,
436
+ 5376
437
+ ],
438
+ "sample_rows": 64,
439
+ "absmax_error": 0.00881594605743885,
440
+ "searched_error": 0.008770952932536602,
441
+ "improvement": 4.499312490224838e-05,
442
+ "clipped_row_fraction": 0.15625,
443
+ "bf16_delta_bytes": 115519416
444
+ },
445
+ {
446
+ "layer": "blocks.7.attn.out_proj",
447
+ "block": 7,
448
+ "kind": "attn.out_proj",
449
+ "shape": [
450
+ 5376,
451
+ 7168
452
+ ],
453
+ "sample_rows": 64,
454
+ "absmax_error": 0.011022337712347507,
455
+ "searched_error": 0.010928143747150898,
456
+ "improvement": 9.41939651966095e-05,
457
+ "clipped_row_fraction": 0.28125,
458
+ "bf16_delta_bytes": 38513592
459
+ },
460
+ {
461
+ "layer": "blocks.7.mlp.fc1",
462
+ "block": 7,
463
+ "kind": "mlp.fc1",
464
+ "shape": [
465
+ 28672,
466
+ 5376
467
+ ],
468
+ "sample_rows": 64,
469
+ "absmax_error": 0.008879266679286957,
470
+ "searched_error": 0.008849864825606346,
471
+ "improvement": 2.9401853680610657e-05,
472
+ "clipped_row_fraction": 0.140625,
473
+ "bf16_delta_bytes": 154025912
474
+ },
475
+ {
476
+ "layer": "blocks.7.mlp.fc2",
477
+ "block": 7,
478
+ "kind": "mlp.fc2",
479
+ "shape": [
480
+ 5376,
481
+ 14336
482
+ ],
483
+ "sample_rows": 64,
484
+ "absmax_error": 0.009405774064362049,
485
+ "searched_error": 0.009109561331570148,
486
+ "improvement": 0.00029621273279190063,
487
+ "clipped_row_fraction": 0.8125,
488
+ "bf16_delta_bytes": 77048760
489
+ },
490
+ {
491
+ "layer": "blocks.8.attn.qkv_proj",
492
+ "block": 8,
493
+ "kind": "attn.qkv_proj",
494
+ "shape": [
495
+ 21504,
496
+ 5376
497
+ ],
498
+ "sample_rows": 64,
499
+ "absmax_error": 0.009008223190903664,
500
+ "searched_error": 0.0089647201821208,
501
+ "improvement": 4.350300878286362e-05,
502
+ "clipped_row_fraction": 0.171875,
503
+ "bf16_delta_bytes": 115519416
504
+ },
505
+ {
506
+ "layer": "blocks.8.attn.out_proj",
507
+ "block": 8,
508
+ "kind": "attn.out_proj",
509
+ "shape": [
510
+ 5376,
511
+ 7168
512
+ ],
513
+ "sample_rows": 64,
514
+ "absmax_error": 0.01042734831571579,
515
+ "searched_error": 0.010339769534766674,
516
+ "improvement": 8.757878094911575e-05,
517
+ "clipped_row_fraction": 0.3125,
518
+ "bf16_delta_bytes": 38513592
519
+ },
520
+ {
521
+ "layer": "blocks.8.mlp.fc1",
522
+ "block": 8,
523
+ "kind": "mlp.fc1",
524
+ "shape": [
525
+ 28672,
526
+ 5376
527
+ ],
528
+ "sample_rows": 64,
529
+ "absmax_error": 0.008798491209745407,
530
+ "searched_error": 0.008746922016143799,
531
+ "improvement": 5.1569193601608276e-05,
532
+ "clipped_row_fraction": 0.1875,
533
+ "bf16_delta_bytes": 154025912
534
+ },
535
+ {
536
+ "layer": "blocks.8.mlp.fc2",
537
+ "block": 8,
538
+ "kind": "mlp.fc2",
539
+ "shape": [
540
+ 5376,
541
+ 14336
542
+ ],
543
+ "sample_rows": 64,
544
+ "absmax_error": 0.009286905638873577,
545
+ "searched_error": 0.009002236649394035,
546
+ "improvement": 0.0002846689894795418,
547
+ "clipped_row_fraction": 0.84375,
548
+ "bf16_delta_bytes": 77048760
549
+ },
550
+ {
551
+ "layer": "blocks.9.attn.qkv_proj",
552
+ "block": 9,
553
+ "kind": "attn.qkv_proj",
554
+ "shape": [
555
+ 21504,
556
+ 5376
557
+ ],
558
+ "sample_rows": 64,
559
+ "absmax_error": 0.008678901940584183,
560
+ "searched_error": 0.008630734868347645,
561
+ "improvement": 4.816707223653793e-05,
562
+ "clipped_row_fraction": 0.1875,
563
+ "bf16_delta_bytes": 115519416
564
+ },
565
+ {
566
+ "layer": "blocks.9.attn.out_proj",
567
+ "block": 9,
568
+ "kind": "attn.out_proj",
569
+ "shape": [
570
+ 5376,
571
+ 7168
572
+ ],
573
+ "sample_rows": 64,
574
+ "absmax_error": 0.010669376701116562,
575
+ "searched_error": 0.010587197728455067,
576
+ "improvement": 8.217897266149521e-05,
577
+ "clipped_row_fraction": 0.265625,
578
+ "bf16_delta_bytes": 38513592
579
+ },
580
+ {
581
+ "layer": "blocks.9.mlp.fc1",
582
+ "block": 9,
583
+ "kind": "mlp.fc1",
584
+ "shape": [
585
+ 28672,
586
+ 5376
587
+ ],
588
+ "sample_rows": 64,
589
+ "absmax_error": 0.008611093275249004,
590
+ "searched_error": 0.008564319461584091,
591
+ "improvement": 4.677381366491318e-05,
592
+ "clipped_row_fraction": 0.15625,
593
+ "bf16_delta_bytes": 154025912
594
+ },
595
+ {
596
+ "layer": "blocks.9.mlp.fc2",
597
+ "block": 9,
598
+ "kind": "mlp.fc2",
599
+ "shape": [
600
+ 5376,
601
+ 14336
602
+ ],
603
+ "sample_rows": 64,
604
+ "absmax_error": 0.009352547116577625,
605
+ "searched_error": 0.009089149534702301,
606
+ "improvement": 0.00026339758187532425,
607
+ "clipped_row_fraction": 0.765625,
608
+ "bf16_delta_bytes": 77048760
609
+ },
610
+ {
611
+ "layer": "blocks.10.attn.qkv_proj",
612
+ "block": 10,
613
+ "kind": "attn.qkv_proj",
614
+ "shape": [
615
+ 21504,
616
+ 5376
617
+ ],
618
+ "sample_rows": 64,
619
+ "absmax_error": 0.00882525835186243,
620
+ "searched_error": 0.008772222325205803,
621
+ "improvement": 5.3036026656627655e-05,
622
+ "clipped_row_fraction": 0.234375,
623
+ "bf16_delta_bytes": 115519416
624
+ },
625
+ {
626
+ "layer": "blocks.10.attn.out_proj",
627
+ "block": 10,
628
+ "kind": "attn.out_proj",
629
+ "shape": [
630
+ 5376,
631
+ 7168
632
+ ],
633
+ "sample_rows": 64,
634
+ "absmax_error": 0.01042264699935913,
635
+ "searched_error": 0.010306568816304207,
636
+ "improvement": 0.00011607818305492401,
637
+ "clipped_row_fraction": 0.375,
638
+ "bf16_delta_bytes": 38513592
639
+ },
640
+ {
641
+ "layer": "blocks.10.mlp.fc1",
642
+ "block": 10,
643
+ "kind": "mlp.fc1",
644
+ "shape": [
645
+ 28672,
646
+ 5376
647
+ ],
648
+ "sample_rows": 64,
649
+ "absmax_error": 0.008798566646873951,
650
+ "searched_error": 0.008739395067095757,
651
+ "improvement": 5.917157977819443e-05,
652
+ "clipped_row_fraction": 0.25,
653
+ "bf16_delta_bytes": 154025912
654
+ },
655
+ {
656
+ "layer": "blocks.10.mlp.fc2",
657
+ "block": 10,
658
+ "kind": "mlp.fc2",
659
+ "shape": [
660
+ 5376,
661
+ 14336
662
+ ],
663
+ "sample_rows": 64,
664
+ "absmax_error": 0.009307919070124626,
665
+ "searched_error": 0.00904966052621603,
666
+ "improvement": 0.00025825854390859604,
667
+ "clipped_row_fraction": 0.765625,
668
+ "bf16_delta_bytes": 77048760
669
+ },
670
+ {
671
+ "layer": "blocks.11.attn.qkv_proj",
672
+ "block": 11,
673
+ "kind": "attn.qkv_proj",
674
+ "shape": [
675
+ 21504,
676
+ 5376
677
+ ],
678
+ "sample_rows": 64,
679
+ "absmax_error": 0.0088732298463583,
680
+ "searched_error": 0.008829551748931408,
681
+ "improvement": 4.367809742689133e-05,
682
+ "clipped_row_fraction": 0.15625,
683
+ "bf16_delta_bytes": 115519416
684
+ },
685
+ {
686
+ "layer": "blocks.11.attn.out_proj",
687
+ "block": 11,
688
+ "kind": "attn.out_proj",
689
+ "shape": [
690
+ 5376,
691
+ 7168
692
+ ],
693
+ "sample_rows": 64,
694
+ "absmax_error": 0.010403656400740147,
695
+ "searched_error": 0.01028230506926775,
696
+ "improvement": 0.00012135133147239685,
697
+ "clipped_row_fraction": 0.359375,
698
+ "bf16_delta_bytes": 38513592
699
+ },
700
+ {
701
+ "layer": "blocks.11.mlp.fc1",
702
+ "block": 11,
703
+ "kind": "mlp.fc1",
704
+ "shape": [
705
+ 28672,
706
+ 5376
707
+ ],
708
+ "sample_rows": 64,
709
+ "absmax_error": 0.008903699927031994,
710
+ "searched_error": 0.008855773136019707,
711
+ "improvement": 4.792679101228714e-05,
712
+ "clipped_row_fraction": 0.1875,
713
+ "bf16_delta_bytes": 154025912
714
+ },
715
+ {
716
+ "layer": "blocks.11.mlp.fc2",
717
+ "block": 11,
718
+ "kind": "mlp.fc2",
719
+ "shape": [
720
+ 5376,
721
+ 14336
722
+ ],
723
+ "sample_rows": 64,
724
+ "absmax_error": 0.009393034502863884,
725
+ "searched_error": 0.009126064367592335,
726
+ "improvement": 0.0002669701352715492,
727
+ "clipped_row_fraction": 0.734375,
728
+ "bf16_delta_bytes": 77048760
729
+ },
730
+ {
731
+ "layer": "blocks.12.attn.qkv_proj",
732
+ "block": 12,
733
+ "kind": "attn.qkv_proj",
734
+ "shape": [
735
+ 21504,
736
+ 5376
737
+ ],
738
+ "sample_rows": 64,
739
+ "absmax_error": 0.008730018511414528,
740
+ "searched_error": 0.008657277561724186,
741
+ "improvement": 7.274094969034195e-05,
742
+ "clipped_row_fraction": 0.234375,
743
+ "bf16_delta_bytes": 115519416
744
+ },
745
+ {
746
+ "layer": "blocks.12.attn.out_proj",
747
+ "block": 12,
748
+ "kind": "attn.out_proj",
749
+ "shape": [
750
+ 5376,
751
+ 7168
752
+ ],
753
+ "sample_rows": 64,
754
+ "absmax_error": 0.010259442962706089,
755
+ "searched_error": 0.010190111584961414,
756
+ "improvement": 6.933137774467468e-05,
757
+ "clipped_row_fraction": 0.296875,
758
+ "bf16_delta_bytes": 38513592
759
+ },
760
+ {
761
+ "layer": "blocks.12.mlp.fc1",
762
+ "block": 12,
763
+ "kind": "mlp.fc1",
764
+ "shape": [
765
+ 28672,
766
+ 5376
767
+ ],
768
+ "sample_rows": 64,
769
+ "absmax_error": 0.008771385997533798,
770
+ "searched_error": 0.00874131079763174,
771
+ "improvement": 3.0075199902057648e-05,
772
+ "clipped_row_fraction": 0.171875,
773
+ "bf16_delta_bytes": 154025912
774
+ },
775
+ {
776
+ "layer": "blocks.12.mlp.fc2",
777
+ "block": 12,
778
+ "kind": "mlp.fc2",
779
+ "shape": [
780
+ 5376,
781
+ 14336
782
+ ],
783
+ "sample_rows": 64,
784
+ "absmax_error": 0.009368893690407276,
785
+ "searched_error": 0.009100206196308136,
786
+ "improvement": 0.00026868749409914017,
787
+ "clipped_row_fraction": 0.75,
788
+ "bf16_delta_bytes": 77048760
789
+ },
790
+ {
791
+ "layer": "blocks.13.attn.qkv_proj",
792
+ "block": 13,
793
+ "kind": "attn.qkv_proj",
794
+ "shape": [
795
+ 21504,
796
+ 5376
797
+ ],
798
+ "sample_rows": 64,
799
+ "absmax_error": 0.008816634304821491,
800
+ "searched_error": 0.008743951097130775,
801
+ "improvement": 7.268320769071579e-05,
802
+ "clipped_row_fraction": 0.25,
803
+ "bf16_delta_bytes": 115519416
804
+ },
805
+ {
806
+ "layer": "blocks.13.attn.out_proj",
807
+ "block": 13,
808
+ "kind": "attn.out_proj",
809
+ "shape": [
810
+ 5376,
811
+ 7168
812
+ ],
813
+ "sample_rows": 64,
814
+ "absmax_error": 0.010227742604911327,
815
+ "searched_error": 0.010140802711248398,
816
+ "improvement": 8.693989366292953e-05,
817
+ "clipped_row_fraction": 0.296875,
818
+ "bf16_delta_bytes": 38513592
819
+ },
820
+ {
821
+ "layer": "blocks.13.mlp.fc1",
822
+ "block": 13,
823
+ "kind": "mlp.fc1",
824
+ "shape": [
825
+ 28672,
826
+ 5376
827
+ ],
828
+ "sample_rows": 64,
829
+ "absmax_error": 0.00874239206314087,
830
+ "searched_error": 0.008691736496984959,
831
+ "improvement": 5.065556615591049e-05,
832
+ "clipped_row_fraction": 0.203125,
833
+ "bf16_delta_bytes": 154025912
834
+ },
835
+ {
836
+ "layer": "blocks.13.mlp.fc2",
837
+ "block": 13,
838
+ "kind": "mlp.fc2",
839
+ "shape": [
840
+ 5376,
841
+ 14336
842
+ ],
843
+ "sample_rows": 64,
844
+ "absmax_error": 0.009334002621471882,
845
+ "searched_error": 0.009064814075827599,
846
+ "improvement": 0.0002691885456442833,
847
+ "clipped_row_fraction": 0.796875,
848
+ "bf16_delta_bytes": 77048760
849
+ },
850
+ {
851
+ "layer": "blocks.14.attn.qkv_proj",
852
+ "block": 14,
853
+ "kind": "attn.qkv_proj",
854
+ "shape": [
855
+ 21504,
856
+ 5376
857
+ ],
858
+ "sample_rows": 64,
859
+ "absmax_error": 0.008730604313313961,
860
+ "searched_error": 0.008698868565261364,
861
+ "improvement": 3.1735748052597046e-05,
862
+ "clipped_row_fraction": 0.109375,
863
+ "bf16_delta_bytes": 115519416
864
+ },
865
+ {
866
+ "layer": "blocks.14.attn.out_proj",
867
+ "block": 14,
868
+ "kind": "attn.out_proj",
869
+ "shape": [
870
+ 5376,
871
+ 7168
872
+ ],
873
+ "sample_rows": 64,
874
+ "absmax_error": 0.010583301074802876,
875
+ "searched_error": 0.010479689575731754,
876
+ "improvement": 0.00010361149907112122,
877
+ "clipped_row_fraction": 0.359375,
878
+ "bf16_delta_bytes": 38513592
879
+ },
880
+ {
881
+ "layer": "blocks.14.mlp.fc1",
882
+ "block": 14,
883
+ "kind": "mlp.fc1",
884
+ "shape": [
885
+ 28672,
886
+ 5376
887
+ ],
888
+ "sample_rows": 64,
889
+ "absmax_error": 0.00874729547649622,
890
+ "searched_error": 0.008711329661309719,
891
+ "improvement": 3.596581518650055e-05,
892
+ "clipped_row_fraction": 0.171875,
893
+ "bf16_delta_bytes": 154025912
894
+ },
895
+ {
896
+ "layer": "blocks.14.mlp.fc2",
897
+ "block": 14,
898
+ "kind": "mlp.fc2",
899
+ "shape": [
900
+ 5376,
901
+ 14336
902
+ ],
903
+ "sample_rows": 64,
904
+ "absmax_error": 0.009479190222918987,
905
+ "searched_error": 0.009177664294838905,
906
+ "improvement": 0.00030152592808008194,
907
+ "clipped_row_fraction": 0.859375,
908
+ "bf16_delta_bytes": 77048760
909
+ },
910
+ {
911
+ "layer": "blocks.15.attn.qkv_proj",
912
+ "block": 15,
913
+ "kind": "attn.qkv_proj",
914
+ "shape": [
915
+ 21504,
916
+ 5376
917
+ ],
918
+ "sample_rows": 64,
919
+ "absmax_error": 0.008874666877090931,
920
+ "searched_error": 0.008819757029414177,
921
+ "improvement": 5.4909847676754e-05,
922
+ "clipped_row_fraction": 0.234375,
923
+ "bf16_delta_bytes": 115519416
924
+ },
925
+ {
926
+ "layer": "blocks.15.attn.out_proj",
927
+ "block": 15,
928
+ "kind": "attn.out_proj",
929
+ "shape": [
930
+ 5376,
931
+ 7168
932
+ ],
933
+ "sample_rows": 64,
934
+ "absmax_error": 0.011003787629306316,
935
+ "searched_error": 0.010917904786765575,
936
+ "improvement": 8.588284254074097e-05,
937
+ "clipped_row_fraction": 0.3125,
938
+ "bf16_delta_bytes": 38513592
939
+ },
940
+ {
941
+ "layer": "blocks.15.mlp.fc1",
942
+ "block": 15,
943
+ "kind": "mlp.fc1",
944
+ "shape": [
945
+ 28672,
946
+ 5376
947
+ ],
948
+ "sample_rows": 64,
949
+ "absmax_error": 0.008837355300784111,
950
+ "searched_error": 0.00877710897475481,
951
+ "improvement": 6.024632602930069e-05,
952
+ "clipped_row_fraction": 0.21875,
953
+ "bf16_delta_bytes": 154025912
954
+ },
955
+ {
956
+ "layer": "blocks.15.mlp.fc2",
957
+ "block": 15,
958
+ "kind": "mlp.fc2",
959
+ "shape": [
960
+ 5376,
961
+ 14336
962
+ ],
963
+ "sample_rows": 64,
964
+ "absmax_error": 0.009284839034080505,
965
+ "searched_error": 0.008988245390355587,
966
+ "improvement": 0.00029659364372491837,
967
+ "clipped_row_fraction": 0.84375,
968
+ "bf16_delta_bytes": 77048760
969
+ },
970
+ {
971
+ "layer": "blocks.16.attn.qkv_proj",
972
+ "block": 16,
973
+ "kind": "attn.qkv_proj",
974
+ "shape": [
975
+ 21504,
976
+ 5376
977
+ ],
978
+ "sample_rows": 64,
979
+ "absmax_error": 0.008748017251491547,
980
+ "searched_error": 0.008695828728377819,
981
+ "improvement": 5.218852311372757e-05,
982
+ "clipped_row_fraction": 0.1875,
983
+ "bf16_delta_bytes": 115519416
984
+ },
985
+ {
986
+ "layer": "blocks.16.attn.out_proj",
987
+ "block": 16,
988
+ "kind": "attn.out_proj",
989
+ "shape": [
990
+ 5376,
991
+ 7168
992
+ ],
993
+ "sample_rows": 64,
994
+ "absmax_error": 0.00988137535750866,
995
+ "searched_error": 0.009816288948059082,
996
+ "improvement": 6.508640944957733e-05,
997
+ "clipped_row_fraction": 0.28125,
998
+ "bf16_delta_bytes": 38513592
999
+ },
1000
+ {
1001
+ "layer": "blocks.16.mlp.fc1",
1002
+ "block": 16,
1003
+ "kind": "mlp.fc1",
1004
+ "shape": [
1005
+ 28672,
1006
+ 5376
1007
+ ],
1008
+ "sample_rows": 64,
1009
+ "absmax_error": 0.008817206136882305,
1010
+ "searched_error": 0.008756496012210846,
1011
+ "improvement": 6.07101246714592e-05,
1012
+ "clipped_row_fraction": 0.25,
1013
+ "bf16_delta_bytes": 154025912
1014
+ },
1015
+ {
1016
+ "layer": "blocks.16.mlp.fc2",
1017
+ "block": 16,
1018
+ "kind": "mlp.fc2",
1019
+ "shape": [
1020
+ 5376,
1021
+ 14336
1022
+ ],
1023
+ "sample_rows": 64,
1024
+ "absmax_error": 0.009356603026390076,
1025
+ "searched_error": 0.009065966121852398,
1026
+ "improvement": 0.00029063690453767776,
1027
+ "clipped_row_fraction": 0.859375,
1028
+ "bf16_delta_bytes": 77048760
1029
+ },
1030
+ {
1031
+ "layer": "blocks.17.attn.qkv_proj",
1032
+ "block": 17,
1033
+ "kind": "attn.qkv_proj",
1034
+ "shape": [
1035
+ 21504,
1036
+ 5376
1037
+ ],
1038
+ "sample_rows": 64,
1039
+ "absmax_error": 0.008703425526618958,
1040
+ "searched_error": 0.008664635941386223,
1041
+ "improvement": 3.878958523273468e-05,
1042
+ "clipped_row_fraction": 0.15625,
1043
+ "bf16_delta_bytes": 115519416
1044
+ },
1045
+ {
1046
+ "layer": "blocks.17.attn.out_proj",
1047
+ "block": 17,
1048
+ "kind": "attn.out_proj",
1049
+ "shape": [
1050
+ 5376,
1051
+ 7168
1052
+ ],
1053
+ "sample_rows": 64,
1054
+ "absmax_error": 0.010421531274914742,
1055
+ "searched_error": 0.01031559333205223,
1056
+ "improvement": 0.00010593794286251068,
1057
+ "clipped_row_fraction": 0.390625,
1058
+ "bf16_delta_bytes": 38513592
1059
+ },
1060
+ {
1061
+ "layer": "blocks.17.mlp.fc1",
1062
+ "block": 17,
1063
+ "kind": "mlp.fc1",
1064
+ "shape": [
1065
+ 28672,
1066
+ 5376
1067
+ ],
1068
+ "sample_rows": 64,
1069
+ "absmax_error": 0.00872599147260189,
1070
+ "searched_error": 0.00865576695650816,
1071
+ "improvement": 7.022451609373093e-05,
1072
+ "clipped_row_fraction": 0.234375,
1073
+ "bf16_delta_bytes": 154025912
1074
+ },
1075
+ {
1076
+ "layer": "blocks.17.mlp.fc2",
1077
+ "block": 17,
1078
+ "kind": "mlp.fc2",
1079
+ "shape": [
1080
+ 5376,
1081
+ 14336
1082
+ ],
1083
+ "sample_rows": 64,
1084
+ "absmax_error": 0.009331951849162579,
1085
+ "searched_error": 0.009054707363247871,
1086
+ "improvement": 0.0002772444859147072,
1087
+ "clipped_row_fraction": 0.828125,
1088
+ "bf16_delta_bytes": 77048760
1089
+ },
1090
+ {
1091
+ "layer": "blocks.18.attn.qkv_proj",
1092
+ "block": 18,
1093
+ "kind": "attn.qkv_proj",
1094
+ "shape": [
1095
+ 21504,
1096
+ 5376
1097
+ ],
1098
+ "sample_rows": 64,
1099
+ "absmax_error": 0.008934194222092628,
1100
+ "searched_error": 0.008866879157721996,
1101
+ "improvement": 6.731506437063217e-05,
1102
+ "clipped_row_fraction": 0.234375,
1103
+ "bf16_delta_bytes": 115519416
1104
+ },
1105
+ {
1106
+ "layer": "blocks.18.attn.out_proj",
1107
+ "block": 18,
1108
+ "kind": "attn.out_proj",
1109
+ "shape": [
1110
+ 5376,
1111
+ 7168
1112
+ ],
1113
+ "sample_rows": 64,
1114
+ "absmax_error": 0.009590333327651024,
1115
+ "searched_error": 0.009500373154878616,
1116
+ "improvement": 8.996017277240753e-05,
1117
+ "clipped_row_fraction": 0.296875,
1118
+ "bf16_delta_bytes": 38513592
1119
+ },
1120
+ {
1121
+ "layer": "blocks.18.mlp.fc1",
1122
+ "block": 18,
1123
+ "kind": "mlp.fc1",
1124
+ "shape": [
1125
+ 28672,
1126
+ 5376
1127
+ ],
1128
+ "sample_rows": 64,
1129
+ "absmax_error": 0.0087647195905447,
1130
+ "searched_error": 0.00870244950056076,
1131
+ "improvement": 6.227008998394012e-05,
1132
+ "clipped_row_fraction": 0.296875,
1133
+ "bf16_delta_bytes": 154025912
1134
+ },
1135
+ {
1136
+ "layer": "blocks.18.mlp.fc2",
1137
+ "block": 18,
1138
+ "kind": "mlp.fc2",
1139
+ "shape": [
1140
+ 5376,
1141
+ 14336
1142
+ ],
1143
+ "sample_rows": 64,
1144
+ "absmax_error": 0.009473690763115883,
1145
+ "searched_error": 0.009212583303451538,
1146
+ "improvement": 0.0002611074596643448,
1147
+ "clipped_row_fraction": 0.75,
1148
+ "bf16_delta_bytes": 77048760
1149
+ },
1150
+ {
1151
+ "layer": "blocks.19.attn.qkv_proj",
1152
+ "block": 19,
1153
+ "kind": "attn.qkv_proj",
1154
+ "shape": [
1155
+ 21504,
1156
+ 5376
1157
+ ],
1158
+ "sample_rows": 64,
1159
+ "absmax_error": 0.00883630383759737,
1160
+ "searched_error": 0.008758751675486565,
1161
+ "improvement": 7.755216211080551e-05,
1162
+ "clipped_row_fraction": 0.265625,
1163
+ "bf16_delta_bytes": 115519416
1164
+ },
1165
+ {
1166
+ "layer": "blocks.19.attn.out_proj",
1167
+ "block": 19,
1168
+ "kind": "attn.out_proj",
1169
+ "shape": [
1170
+ 5376,
1171
+ 7168
1172
+ ],
1173
+ "sample_rows": 64,
1174
+ "absmax_error": 0.010613351128995419,
1175
+ "searched_error": 0.010535142384469509,
1176
+ "improvement": 7.820874452590942e-05,
1177
+ "clipped_row_fraction": 0.296875,
1178
+ "bf16_delta_bytes": 38513592
1179
+ },
1180
+ {
1181
+ "layer": "blocks.19.mlp.fc1",
1182
+ "block": 19,
1183
+ "kind": "mlp.fc1",
1184
+ "shape": [
1185
+ 28672,
1186
+ 5376
1187
+ ],
1188
+ "sample_rows": 64,
1189
+ "absmax_error": 0.008808483369648457,
1190
+ "searched_error": 0.00879175029695034,
1191
+ "improvement": 1.6733072698116302e-05,
1192
+ "clipped_row_fraction": 0.09375,
1193
+ "bf16_delta_bytes": 154025912
1194
+ },
1195
+ {
1196
+ "layer": "blocks.19.mlp.fc2",
1197
+ "block": 19,
1198
+ "kind": "mlp.fc2",
1199
+ "shape": [
1200
+ 5376,
1201
+ 14336
1202
+ ],
1203
+ "sample_rows": 64,
1204
+ "absmax_error": 0.009467557072639465,
1205
+ "searched_error": 0.009168003685772419,
1206
+ "improvement": 0.00029955338686704636,
1207
+ "clipped_row_fraction": 0.859375,
1208
+ "bf16_delta_bytes": 77048760
1209
+ },
1210
+ {
1211
+ "layer": "blocks.20.attn.qkv_proj",
1212
+ "block": 20,
1213
+ "kind": "attn.qkv_proj",
1214
+ "shape": [
1215
+ 21504,
1216
+ 5376
1217
+ ],
1218
+ "sample_rows": 64,
1219
+ "absmax_error": 0.008711456321179867,
1220
+ "searched_error": 0.00862672459334135,
1221
+ "improvement": 8.473172783851624e-05,
1222
+ "clipped_row_fraction": 0.3125,
1223
+ "bf16_delta_bytes": 115519416
1224
+ },
1225
+ {
1226
+ "layer": "blocks.20.attn.out_proj",
1227
+ "block": 20,
1228
+ "kind": "attn.out_proj",
1229
+ "shape": [
1230
+ 5376,
1231
+ 7168
1232
+ ],
1233
+ "sample_rows": 64,
1234
+ "absmax_error": 0.010097628459334373,
1235
+ "searched_error": 0.010016809217631817,
1236
+ "improvement": 8.081924170255661e-05,
1237
+ "clipped_row_fraction": 0.328125,
1238
+ "bf16_delta_bytes": 38513592
1239
+ },
1240
+ {
1241
+ "layer": "blocks.20.mlp.fc1",
1242
+ "block": 20,
1243
+ "kind": "mlp.fc1",
1244
+ "shape": [
1245
+ 28672,
1246
+ 5376
1247
+ ],
1248
+ "sample_rows": 64,
1249
+ "absmax_error": 0.008967501111328602,
1250
+ "searched_error": 0.008894388563930988,
1251
+ "improvement": 7.311254739761353e-05,
1252
+ "clipped_row_fraction": 0.296875,
1253
+ "bf16_delta_bytes": 154025912
1254
+ },
1255
+ {
1256
+ "layer": "blocks.20.mlp.fc2",
1257
+ "block": 20,
1258
+ "kind": "mlp.fc2",
1259
+ "shape": [
1260
+ 5376,
1261
+ 14336
1262
+ ],
1263
+ "sample_rows": 64,
1264
+ "absmax_error": 0.009428652003407478,
1265
+ "searched_error": 0.00914465170353651,
1266
+ "improvement": 0.00028400029987096786,
1267
+ "clipped_row_fraction": 0.796875,
1268
+ "bf16_delta_bytes": 77048760
1269
+ },
1270
+ {
1271
+ "layer": "blocks.21.attn.qkv_proj",
1272
+ "block": 21,
1273
+ "kind": "attn.qkv_proj",
1274
+ "shape": [
1275
+ 21504,
1276
+ 5376
1277
+ ],
1278
+ "sample_rows": 64,
1279
+ "absmax_error": 0.0088729839771986,
1280
+ "searched_error": 0.00883404165506363,
1281
+ "improvement": 3.894232213497162e-05,
1282
+ "clipped_row_fraction": 0.15625,
1283
+ "bf16_delta_bytes": 115519416
1284
+ },
1285
+ {
1286
+ "layer": "blocks.21.attn.out_proj",
1287
+ "block": 21,
1288
+ "kind": "attn.out_proj",
1289
+ "shape": [
1290
+ 5376,
1291
+ 7168
1292
+ ],
1293
+ "sample_rows": 64,
1294
+ "absmax_error": 0.009565706364810467,
1295
+ "searched_error": 0.009480264037847519,
1296
+ "improvement": 8.544232696294785e-05,
1297
+ "clipped_row_fraction": 0.3125,
1298
+ "bf16_delta_bytes": 38513592
1299
+ },
1300
+ {
1301
+ "layer": "blocks.21.mlp.fc1",
1302
+ "block": 21,
1303
+ "kind": "mlp.fc1",
1304
+ "shape": [
1305
+ 28672,
1306
+ 5376
1307
+ ],
1308
+ "sample_rows": 64,
1309
+ "absmax_error": 0.00895257480442524,
1310
+ "searched_error": 0.00890747457742691,
1311
+ "improvement": 4.510022699832916e-05,
1312
+ "clipped_row_fraction": 0.15625,
1313
+ "bf16_delta_bytes": 154025912
1314
+ },
1315
+ {
1316
+ "layer": "blocks.21.mlp.fc2",
1317
+ "block": 21,
1318
+ "kind": "mlp.fc2",
1319
+ "shape": [
1320
+ 5376,
1321
+ 14336
1322
+ ],
1323
+ "sample_rows": 64,
1324
+ "absmax_error": 0.009270595386624336,
1325
+ "searched_error": 0.008980251848697662,
1326
+ "improvement": 0.0002903435379266739,
1327
+ "clipped_row_fraction": 0.84375,
1328
+ "bf16_delta_bytes": 77048760
1329
+ },
1330
+ {
1331
+ "layer": "blocks.22.attn.qkv_proj",
1332
+ "block": 22,
1333
+ "kind": "attn.qkv_proj",
1334
+ "shape": [
1335
+ 21504,
1336
+ 5376
1337
+ ],
1338
+ "sample_rows": 64,
1339
+ "absmax_error": 0.00894355308264494,
1340
+ "searched_error": 0.00891187135130167,
1341
+ "improvement": 3.168173134326935e-05,
1342
+ "clipped_row_fraction": 0.140625,
1343
+ "bf16_delta_bytes": 115519416
1344
+ },
1345
+ {
1346
+ "layer": "blocks.22.attn.out_proj",
1347
+ "block": 22,
1348
+ "kind": "attn.out_proj",
1349
+ "shape": [
1350
+ 5376,
1351
+ 7168
1352
+ ],
1353
+ "sample_rows": 64,
1354
+ "absmax_error": 0.00975025724619627,
1355
+ "searched_error": 0.009666379541158676,
1356
+ "improvement": 8.387770503759384e-05,
1357
+ "clipped_row_fraction": 0.28125,
1358
+ "bf16_delta_bytes": 38513592
1359
+ },
1360
+ {
1361
+ "layer": "blocks.22.mlp.fc1",
1362
+ "block": 22,
1363
+ "kind": "mlp.fc1",
1364
+ "shape": [
1365
+ 28672,
1366
+ 5376
1367
+ ],
1368
+ "sample_rows": 64,
1369
+ "absmax_error": 0.00883104931563139,
1370
+ "searched_error": 0.008784261532127857,
1371
+ "improvement": 4.678778350353241e-05,
1372
+ "clipped_row_fraction": 0.171875,
1373
+ "bf16_delta_bytes": 154025912
1374
+ },
1375
+ {
1376
+ "layer": "blocks.22.mlp.fc2",
1377
+ "block": 22,
1378
+ "kind": "mlp.fc2",
1379
+ "shape": [
1380
+ 5376,
1381
+ 14336
1382
+ ],
1383
+ "sample_rows": 64,
1384
+ "absmax_error": 0.009395661763846874,
1385
+ "searched_error": 0.009091234765946865,
1386
+ "improvement": 0.00030442699790000916,
1387
+ "clipped_row_fraction": 0.875,
1388
+ "bf16_delta_bytes": 77048760
1389
+ },
1390
+ {
1391
+ "layer": "blocks.23.attn.qkv_proj",
1392
+ "block": 23,
1393
+ "kind": "attn.qkv_proj",
1394
+ "shape": [
1395
+ 21504,
1396
+ 5376
1397
+ ],
1398
+ "sample_rows": 64,
1399
+ "absmax_error": 0.00885548535734415,
1400
+ "searched_error": 0.008821829222142696,
1401
+ "improvement": 3.365613520145416e-05,
1402
+ "clipped_row_fraction": 0.125,
1403
+ "bf16_delta_bytes": 115519416
1404
+ },
1405
+ {
1406
+ "layer": "blocks.23.attn.out_proj",
1407
+ "block": 23,
1408
+ "kind": "attn.out_proj",
1409
+ "shape": [
1410
+ 5376,
1411
+ 7168
1412
+ ],
1413
+ "sample_rows": 64,
1414
+ "absmax_error": 0.009813772514462471,
1415
+ "searched_error": 0.009709117002785206,
1416
+ "improvement": 0.00010465551167726517,
1417
+ "clipped_row_fraction": 0.3125,
1418
+ "bf16_delta_bytes": 38513592
1419
+ },
1420
+ {
1421
+ "layer": "blocks.23.mlp.fc1",
1422
+ "block": 23,
1423
+ "kind": "mlp.fc1",
1424
+ "shape": [
1425
+ 28672,
1426
+ 5376
1427
+ ],
1428
+ "sample_rows": 64,
1429
+ "absmax_error": 0.008615493774414062,
1430
+ "searched_error": 0.008556989952921867,
1431
+ "improvement": 5.850382149219513e-05,
1432
+ "clipped_row_fraction": 0.25,
1433
+ "bf16_delta_bytes": 154025912
1434
+ },
1435
+ {
1436
+ "layer": "blocks.23.mlp.fc2",
1437
+ "block": 23,
1438
+ "kind": "mlp.fc2",
1439
+ "shape": [
1440
+ 5376,
1441
+ 14336
1442
+ ],
1443
+ "sample_rows": 64,
1444
+ "absmax_error": 0.00936894677579403,
1445
+ "searched_error": 0.009070239961147308,
1446
+ "improvement": 0.0002987068146467209,
1447
+ "clipped_row_fraction": 0.859375,
1448
+ "bf16_delta_bytes": 77048760
1449
+ },
1450
+ {
1451
+ "layer": "blocks.24.attn.qkv_proj",
1452
+ "block": 24,
1453
+ "kind": "attn.qkv_proj",
1454
+ "shape": [
1455
+ 21504,
1456
+ 5376
1457
+ ],
1458
+ "sample_rows": 64,
1459
+ "absmax_error": 0.008828956633806229,
1460
+ "searched_error": 0.008763122372329235,
1461
+ "improvement": 6.583426147699356e-05,
1462
+ "clipped_row_fraction": 0.25,
1463
+ "bf16_delta_bytes": 115519416
1464
+ },
1465
+ {
1466
+ "layer": "blocks.24.attn.out_proj",
1467
+ "block": 24,
1468
+ "kind": "attn.out_proj",
1469
+ "shape": [
1470
+ 5376,
1471
+ 7168
1472
+ ],
1473
+ "sample_rows": 64,
1474
+ "absmax_error": 0.00980150606483221,
1475
+ "searched_error": 0.009715952910482883,
1476
+ "improvement": 8.555315434932709e-05,
1477
+ "clipped_row_fraction": 0.296875,
1478
+ "bf16_delta_bytes": 38513592
1479
+ },
1480
+ {
1481
+ "layer": "blocks.24.mlp.fc1",
1482
+ "block": 24,
1483
+ "kind": "mlp.fc1",
1484
+ "shape": [
1485
+ 28672,
1486
+ 5376
1487
+ ],
1488
+ "sample_rows": 64,
1489
+ "absmax_error": 0.008843949995934963,
1490
+ "searched_error": 0.008792220614850521,
1491
+ "improvement": 5.172938108444214e-05,
1492
+ "clipped_row_fraction": 0.21875,
1493
+ "bf16_delta_bytes": 154025912
1494
+ },
1495
+ {
1496
+ "layer": "blocks.24.mlp.fc2",
1497
+ "block": 24,
1498
+ "kind": "mlp.fc2",
1499
+ "shape": [
1500
+ 5376,
1501
+ 14336
1502
+ ],
1503
+ "sample_rows": 64,
1504
+ "absmax_error": 0.009455005638301373,
1505
+ "searched_error": 0.009169526398181915,
1506
+ "improvement": 0.00028547924011945724,
1507
+ "clipped_row_fraction": 0.8125,
1508
+ "bf16_delta_bytes": 77048760
1509
+ },
1510
+ {
1511
+ "layer": "blocks.25.attn.qkv_proj",
1512
+ "block": 25,
1513
+ "kind": "attn.qkv_proj",
1514
+ "shape": [
1515
+ 21504,
1516
+ 5376
1517
+ ],
1518
+ "sample_rows": 64,
1519
+ "absmax_error": 0.00900735892355442,
1520
+ "searched_error": 0.008946533314883709,
1521
+ "improvement": 6.082560867071152e-05,
1522
+ "clipped_row_fraction": 0.25,
1523
+ "bf16_delta_bytes": 115519416
1524
+ },
1525
+ {
1526
+ "layer": "blocks.25.attn.out_proj",
1527
+ "block": 25,
1528
+ "kind": "attn.out_proj",
1529
+ "shape": [
1530
+ 5376,
1531
+ 7168
1532
+ ],
1533
+ "sample_rows": 64,
1534
+ "absmax_error": 0.009812835603952408,
1535
+ "searched_error": 0.00972868874669075,
1536
+ "improvement": 8.414685726165771e-05,
1537
+ "clipped_row_fraction": 0.265625,
1538
+ "bf16_delta_bytes": 38513592
1539
+ },
1540
+ {
1541
+ "layer": "blocks.25.mlp.fc1",
1542
+ "block": 25,
1543
+ "kind": "mlp.fc1",
1544
+ "shape": [
1545
+ 28672,
1546
+ 5376
1547
+ ],
1548
+ "sample_rows": 64,
1549
+ "absmax_error": 0.008857025764882565,
1550
+ "searched_error": 0.008804724551737309,
1551
+ "improvement": 5.230121314525604e-05,
1552
+ "clipped_row_fraction": 0.1875,
1553
+ "bf16_delta_bytes": 154025912
1554
+ },
1555
+ {
1556
+ "layer": "blocks.25.mlp.fc2",
1557
+ "block": 25,
1558
+ "kind": "mlp.fc2",
1559
+ "shape": [
1560
+ 5376,
1561
+ 14336
1562
+ ],
1563
+ "sample_rows": 64,
1564
+ "absmax_error": 0.009359694086015224,
1565
+ "searched_error": 0.009061169810593128,
1566
+ "improvement": 0.00029852427542209625,
1567
+ "clipped_row_fraction": 0.875,
1568
+ "bf16_delta_bytes": 77048760
1569
+ },
1570
+ {
1571
+ "layer": "blocks.26.attn.qkv_proj",
1572
+ "block": 26,
1573
+ "kind": "attn.qkv_proj",
1574
+ "shape": [
1575
+ 21504,
1576
+ 5376
1577
+ ],
1578
+ "sample_rows": 64,
1579
+ "absmax_error": 0.008898529224097729,
1580
+ "searched_error": 0.008845372125506401,
1581
+ "improvement": 5.315709859132767e-05,
1582
+ "clipped_row_fraction": 0.1875,
1583
+ "bf16_delta_bytes": 115519416
1584
+ },
1585
+ {
1586
+ "layer": "blocks.26.attn.out_proj",
1587
+ "block": 26,
1588
+ "kind": "attn.out_proj",
1589
+ "shape": [
1590
+ 5376,
1591
+ 7168
1592
+ ],
1593
+ "sample_rows": 64,
1594
+ "absmax_error": 0.010012996383011341,
1595
+ "searched_error": 0.009899193421006203,
1596
+ "improvement": 0.0001138029620051384,
1597
+ "clipped_row_fraction": 0.359375,
1598
+ "bf16_delta_bytes": 38513592
1599
+ },
1600
+ {
1601
+ "layer": "blocks.26.mlp.fc1",
1602
+ "block": 26,
1603
+ "kind": "mlp.fc1",
1604
+ "shape": [
1605
+ 28672,
1606
+ 5376
1607
+ ],
1608
+ "sample_rows": 64,
1609
+ "absmax_error": 0.008813909254968166,
1610
+ "searched_error": 0.008754000067710876,
1611
+ "improvement": 5.9909187257289886e-05,
1612
+ "clipped_row_fraction": 0.234375,
1613
+ "bf16_delta_bytes": 154025912
1614
+ },
1615
+ {
1616
+ "layer": "blocks.26.mlp.fc2",
1617
+ "block": 26,
1618
+ "kind": "mlp.fc2",
1619
+ "shape": [
1620
+ 5376,
1621
+ 14336
1622
+ ],
1623
+ "sample_rows": 64,
1624
+ "absmax_error": 0.009220638312399387,
1625
+ "searched_error": 0.008964918553829193,
1626
+ "improvement": 0.00025571975857019424,
1627
+ "clipped_row_fraction": 0.765625,
1628
+ "bf16_delta_bytes": 77048760
1629
+ },
1630
+ {
1631
+ "layer": "blocks.27.attn.qkv_proj",
1632
+ "block": 27,
1633
+ "kind": "attn.qkv_proj",
1634
+ "shape": [
1635
+ 21504,
1636
+ 5376
1637
+ ],
1638
+ "sample_rows": 64,
1639
+ "absmax_error": 0.008921593427658081,
1640
+ "searched_error": 0.008879701606929302,
1641
+ "improvement": 4.189182072877884e-05,
1642
+ "clipped_row_fraction": 0.171875,
1643
+ "bf16_delta_bytes": 115519416
1644
+ },
1645
+ {
1646
+ "layer": "blocks.27.attn.out_proj",
1647
+ "block": 27,
1648
+ "kind": "attn.out_proj",
1649
+ "shape": [
1650
+ 5376,
1651
+ 7168
1652
+ ],
1653
+ "sample_rows": 64,
1654
+ "absmax_error": 0.010251954197883606,
1655
+ "searched_error": 0.01017802581191063,
1656
+ "improvement": 7.392838597297668e-05,
1657
+ "clipped_row_fraction": 0.28125,
1658
+ "bf16_delta_bytes": 38513592
1659
+ },
1660
+ {
1661
+ "layer": "blocks.27.mlp.fc1",
1662
+ "block": 27,
1663
+ "kind": "mlp.fc1",
1664
+ "shape": [
1665
+ 28672,
1666
+ 5376
1667
+ ],
1668
+ "sample_rows": 64,
1669
+ "absmax_error": 0.008891111239790916,
1670
+ "searched_error": 0.008837737143039703,
1671
+ "improvement": 5.3374096751213074e-05,
1672
+ "clipped_row_fraction": 0.1875,
1673
+ "bf16_delta_bytes": 154025912
1674
+ },
1675
+ {
1676
+ "layer": "blocks.27.mlp.fc2",
1677
+ "block": 27,
1678
+ "kind": "mlp.fc2",
1679
+ "shape": [
1680
+ 5376,
1681
+ 14336
1682
+ ],
1683
+ "sample_rows": 64,
1684
+ "absmax_error": 0.009307934902608395,
1685
+ "searched_error": 0.009034196846187115,
1686
+ "improvement": 0.0002737380564212799,
1687
+ "clipped_row_fraction": 0.78125,
1688
+ "bf16_delta_bytes": 77048760
1689
+ },
1690
+ {
1691
+ "layer": "blocks.28.attn.qkv_proj",
1692
+ "block": 28,
1693
+ "kind": "attn.qkv_proj",
1694
+ "shape": [
1695
+ 21504,
1696
+ 5376
1697
+ ],
1698
+ "sample_rows": 64,
1699
+ "absmax_error": 0.008664083667099476,
1700
+ "searched_error": 0.008631406351923943,
1701
+ "improvement": 3.2677315175533295e-05,
1702
+ "clipped_row_fraction": 0.15625,
1703
+ "bf16_delta_bytes": 115519416
1704
+ },
1705
+ {
1706
+ "layer": "blocks.28.attn.out_proj",
1707
+ "block": 28,
1708
+ "kind": "attn.out_proj",
1709
+ "shape": [
1710
+ 5376,
1711
+ 7168
1712
+ ],
1713
+ "sample_rows": 64,
1714
+ "absmax_error": 0.010012027807533741,
1715
+ "searched_error": 0.009910239838063717,
1716
+ "improvement": 0.00010178796947002411,
1717
+ "clipped_row_fraction": 0.3125,
1718
+ "bf16_delta_bytes": 38513592
1719
+ },
1720
+ {
1721
+ "layer": "blocks.28.mlp.fc1",
1722
+ "block": 28,
1723
+ "kind": "mlp.fc1",
1724
+ "shape": [
1725
+ 28672,
1726
+ 5376
1727
+ ],
1728
+ "sample_rows": 64,
1729
+ "absmax_error": 0.00868856068700552,
1730
+ "searched_error": 0.008628825657069683,
1731
+ "improvement": 5.973502993583679e-05,
1732
+ "clipped_row_fraction": 0.234375,
1733
+ "bf16_delta_bytes": 154025912
1734
+ },
1735
+ {
1736
+ "layer": "blocks.28.mlp.fc2",
1737
+ "block": 28,
1738
+ "kind": "mlp.fc2",
1739
+ "shape": [
1740
+ 5376,
1741
+ 14336
1742
+ ],
1743
+ "sample_rows": 64,
1744
+ "absmax_error": 0.009298889897763729,
1745
+ "searched_error": 0.00900914054363966,
1746
+ "improvement": 0.0002897493541240692,
1747
+ "clipped_row_fraction": 0.828125,
1748
+ "bf16_delta_bytes": 77048760
1749
+ },
1750
+ {
1751
+ "layer": "blocks.29.attn.qkv_proj",
1752
+ "block": 29,
1753
+ "kind": "attn.qkv_proj",
1754
+ "shape": [
1755
+ 21504,
1756
+ 5376
1757
+ ],
1758
+ "sample_rows": 64,
1759
+ "absmax_error": 0.008712724782526493,
1760
+ "searched_error": 0.008658976294100285,
1761
+ "improvement": 5.3748488426208496e-05,
1762
+ "clipped_row_fraction": 0.21875,
1763
+ "bf16_delta_bytes": 115519416
1764
+ },
1765
+ {
1766
+ "layer": "blocks.29.attn.out_proj",
1767
+ "block": 29,
1768
+ "kind": "attn.out_proj",
1769
+ "shape": [
1770
+ 5376,
1771
+ 7168
1772
+ ],
1773
+ "sample_rows": 64,
1774
+ "absmax_error": 0.009863519109785557,
1775
+ "searched_error": 0.009750662371516228,
1776
+ "improvement": 0.00011285673826932907,
1777
+ "clipped_row_fraction": 0.375,
1778
+ "bf16_delta_bytes": 38513592
1779
+ },
1780
+ {
1781
+ "layer": "blocks.29.mlp.fc1",
1782
+ "block": 29,
1783
+ "kind": "mlp.fc1",
1784
+ "shape": [
1785
+ 28672,
1786
+ 5376
1787
+ ],
1788
+ "sample_rows": 64,
1789
+ "absmax_error": 0.008687011897563934,
1790
+ "searched_error": 0.00863416213542223,
1791
+ "improvement": 5.284976214170456e-05,
1792
+ "clipped_row_fraction": 0.1875,
1793
+ "bf16_delta_bytes": 154025912
1794
+ },
1795
+ {
1796
+ "layer": "blocks.29.mlp.fc2",
1797
+ "block": 29,
1798
+ "kind": "mlp.fc2",
1799
+ "shape": [
1800
+ 5376,
1801
+ 14336
1802
+ ],
1803
+ "sample_rows": 64,
1804
+ "absmax_error": 0.00952060241252184,
1805
+ "searched_error": 0.009227273054420948,
1806
+ "improvement": 0.0002933293581008911,
1807
+ "clipped_row_fraction": 0.84375,
1808
+ "bf16_delta_bytes": 77048760
1809
+ },
1810
+ {
1811
+ "layer": "blocks.30.attn.qkv_proj",
1812
+ "block": 30,
1813
+ "kind": "attn.qkv_proj",
1814
+ "shape": [
1815
+ 21504,
1816
+ 5376
1817
+ ],
1818
+ "sample_rows": 64,
1819
+ "absmax_error": 0.008769126608967781,
1820
+ "searched_error": 0.008718494325876236,
1821
+ "improvement": 5.0632283091545105e-05,
1822
+ "clipped_row_fraction": 0.15625,
1823
+ "bf16_delta_bytes": 115519416
1824
+ },
1825
+ {
1826
+ "layer": "blocks.30.attn.out_proj",
1827
+ "block": 30,
1828
+ "kind": "attn.out_proj",
1829
+ "shape": [
1830
+ 5376,
1831
+ 7168
1832
+ ],
1833
+ "sample_rows": 64,
1834
+ "absmax_error": 0.00980446208268404,
1835
+ "searched_error": 0.009715707041323185,
1836
+ "improvement": 8.87550413608551e-05,
1837
+ "clipped_row_fraction": 0.328125,
1838
+ "bf16_delta_bytes": 38513592
1839
+ },
1840
+ {
1841
+ "layer": "blocks.30.mlp.fc1",
1842
+ "block": 30,
1843
+ "kind": "mlp.fc1",
1844
+ "shape": [
1845
+ 28672,
1846
+ 5376
1847
+ ],
1848
+ "sample_rows": 64,
1849
+ "absmax_error": 0.008770755492150784,
1850
+ "searched_error": 0.008700021542608738,
1851
+ "improvement": 7.07339495420456e-05,
1852
+ "clipped_row_fraction": 0.296875,
1853
+ "bf16_delta_bytes": 154025912
1854
+ },
1855
+ {
1856
+ "layer": "blocks.30.mlp.fc2",
1857
+ "block": 30,
1858
+ "kind": "mlp.fc2",
1859
+ "shape": [
1860
+ 5376,
1861
+ 14336
1862
+ ],
1863
+ "sample_rows": 64,
1864
+ "absmax_error": 0.00926683098077774,
1865
+ "searched_error": 0.008997678756713867,
1866
+ "improvement": 0.0002691522240638733,
1867
+ "clipped_row_fraction": 0.765625,
1868
+ "bf16_delta_bytes": 77048760
1869
+ },
1870
+ {
1871
+ "layer": "blocks.31.attn.qkv_proj",
1872
+ "block": 31,
1873
+ "kind": "attn.qkv_proj",
1874
+ "shape": [
1875
+ 21504,
1876
+ 5376
1877
+ ],
1878
+ "sample_rows": 64,
1879
+ "absmax_error": 0.008715620264410973,
1880
+ "searched_error": 0.008677283301949501,
1881
+ "improvement": 3.833696246147156e-05,
1882
+ "clipped_row_fraction": 0.140625,
1883
+ "bf16_delta_bytes": 115519416
1884
+ },
1885
+ {
1886
+ "layer": "blocks.31.attn.out_proj",
1887
+ "block": 31,
1888
+ "kind": "attn.out_proj",
1889
+ "shape": [
1890
+ 5376,
1891
+ 7168
1892
+ ],
1893
+ "sample_rows": 64,
1894
+ "absmax_error": 0.010006719268858433,
1895
+ "searched_error": 0.009893168695271015,
1896
+ "improvement": 0.0001135505735874176,
1897
+ "clipped_row_fraction": 0.375,
1898
+ "bf16_delta_bytes": 38513592
1899
+ },
1900
+ {
1901
+ "layer": "blocks.31.mlp.fc1",
1902
+ "block": 31,
1903
+ "kind": "mlp.fc1",
1904
+ "shape": [
1905
+ 28672,
1906
+ 5376
1907
+ ],
1908
+ "sample_rows": 64,
1909
+ "absmax_error": 0.008751807734370232,
1910
+ "searched_error": 0.008717701770365238,
1911
+ "improvement": 3.410596400499344e-05,
1912
+ "clipped_row_fraction": 0.15625,
1913
+ "bf16_delta_bytes": 154025912
1914
+ },
1915
+ {
1916
+ "layer": "blocks.31.mlp.fc2",
1917
+ "block": 31,
1918
+ "kind": "mlp.fc2",
1919
+ "shape": [
1920
+ 5376,
1921
+ 14336
1922
+ ],
1923
+ "sample_rows": 64,
1924
+ "absmax_error": 0.009457345120608807,
1925
+ "searched_error": 0.009170906618237495,
1926
+ "improvement": 0.0002864385023713112,
1927
+ "clipped_row_fraction": 0.8125,
1928
+ "bf16_delta_bytes": 77048760
1929
+ },
1930
+ {
1931
+ "layer": "blocks.32.attn.qkv_proj",
1932
+ "block": 32,
1933
+ "kind": "attn.qkv_proj",
1934
+ "shape": [
1935
+ 21504,
1936
+ 5376
1937
+ ],
1938
+ "sample_rows": 64,
1939
+ "absmax_error": 0.008660205639898777,
1940
+ "searched_error": 0.008621814660727978,
1941
+ "improvement": 3.8390979170799255e-05,
1942
+ "clipped_row_fraction": 0.140625,
1943
+ "bf16_delta_bytes": 115519416
1944
+ },
1945
+ {
1946
+ "layer": "blocks.32.attn.out_proj",
1947
+ "block": 32,
1948
+ "kind": "attn.out_proj",
1949
+ "shape": [
1950
+ 5376,
1951
+ 7168
1952
+ ],
1953
+ "sample_rows": 64,
1954
+ "absmax_error": 0.009854556061327457,
1955
+ "searched_error": 0.009778095409274101,
1956
+ "improvement": 7.646065205335617e-05,
1957
+ "clipped_row_fraction": 0.28125,
1958
+ "bf16_delta_bytes": 38513592
1959
+ },
1960
+ {
1961
+ "layer": "blocks.32.mlp.fc1",
1962
+ "block": 32,
1963
+ "kind": "mlp.fc1",
1964
+ "shape": [
1965
+ 28672,
1966
+ 5376
1967
+ ],
1968
+ "sample_rows": 64,
1969
+ "absmax_error": 0.008708640933036804,
1970
+ "searched_error": 0.00866017211228609,
1971
+ "improvement": 4.846882075071335e-05,
1972
+ "clipped_row_fraction": 0.171875,
1973
+ "bf16_delta_bytes": 154025912
1974
+ },
1975
+ {
1976
+ "layer": "blocks.32.mlp.fc2",
1977
+ "block": 32,
1978
+ "kind": "mlp.fc2",
1979
+ "shape": [
1980
+ 5376,
1981
+ 14336
1982
+ ],
1983
+ "sample_rows": 64,
1984
+ "absmax_error": 0.009446533396840096,
1985
+ "searched_error": 0.009172597900032997,
1986
+ "improvement": 0.0002739354968070984,
1987
+ "clipped_row_fraction": 0.78125,
1988
+ "bf16_delta_bytes": 77048760
1989
+ },
1990
+ {
1991
+ "layer": "blocks.33.attn.qkv_proj",
1992
+ "block": 33,
1993
+ "kind": "attn.qkv_proj",
1994
+ "shape": [
1995
+ 21504,
1996
+ 5376
1997
+ ],
1998
+ "sample_rows": 64,
1999
+ "absmax_error": 0.008753178641200066,
2000
+ "searched_error": 0.00870854314416647,
2001
+ "improvement": 4.463549703359604e-05,
2002
+ "clipped_row_fraction": 0.1875,
2003
+ "bf16_delta_bytes": 115519416
2004
+ },
2005
+ {
2006
+ "layer": "blocks.33.attn.out_proj",
2007
+ "block": 33,
2008
+ "kind": "attn.out_proj",
2009
+ "shape": [
2010
+ 5376,
2011
+ 7168
2012
+ ],
2013
+ "sample_rows": 64,
2014
+ "absmax_error": 0.009544803760945797,
2015
+ "searched_error": 0.009460123255848885,
2016
+ "improvement": 8.468050509691238e-05,
2017
+ "clipped_row_fraction": 0.3125,
2018
+ "bf16_delta_bytes": 38513592
2019
+ },
2020
+ {
2021
+ "layer": "blocks.33.mlp.fc1",
2022
+ "block": 33,
2023
+ "kind": "mlp.fc1",
2024
+ "shape": [
2025
+ 28672,
2026
+ 5376
2027
+ ],
2028
+ "sample_rows": 64,
2029
+ "absmax_error": 0.00878031924366951,
2030
+ "searched_error": 0.008725587278604507,
2031
+ "improvement": 5.473196506500244e-05,
2032
+ "clipped_row_fraction": 0.171875,
2033
+ "bf16_delta_bytes": 154025912
2034
+ },
2035
+ {
2036
+ "layer": "blocks.33.mlp.fc2",
2037
+ "block": 33,
2038
+ "kind": "mlp.fc2",
2039
+ "shape": [
2040
+ 5376,
2041
+ 14336
2042
+ ],
2043
+ "sample_rows": 64,
2044
+ "absmax_error": 0.009336614981293678,
2045
+ "searched_error": 0.009038848802447319,
2046
+ "improvement": 0.00029776617884635925,
2047
+ "clipped_row_fraction": 0.859375,
2048
+ "bf16_delta_bytes": 77048760
2049
+ },
2050
+ {
2051
+ "layer": "blocks.34.attn.qkv_proj",
2052
+ "block": 34,
2053
+ "kind": "attn.qkv_proj",
2054
+ "shape": [
2055
+ 21504,
2056
+ 5376
2057
+ ],
2058
+ "sample_rows": 64,
2059
+ "absmax_error": 0.00897362269461155,
2060
+ "searched_error": 0.008926293812692165,
2061
+ "improvement": 4.7328881919384e-05,
2062
+ "clipped_row_fraction": 0.1875,
2063
+ "bf16_delta_bytes": 115519416
2064
+ },
2065
+ {
2066
+ "layer": "blocks.34.attn.out_proj",
2067
+ "block": 34,
2068
+ "kind": "attn.out_proj",
2069
+ "shape": [
2070
+ 5376,
2071
+ 7168
2072
+ ],
2073
+ "sample_rows": 64,
2074
+ "absmax_error": 0.009532013908028603,
2075
+ "searched_error": 0.009444000199437141,
2076
+ "improvement": 8.801370859146118e-05,
2077
+ "clipped_row_fraction": 0.34375,
2078
+ "bf16_delta_bytes": 38513592
2079
+ },
2080
+ {
2081
+ "layer": "blocks.34.mlp.fc1",
2082
+ "block": 34,
2083
+ "kind": "mlp.fc1",
2084
+ "shape": [
2085
+ 28672,
2086
+ 5376
2087
+ ],
2088
+ "sample_rows": 64,
2089
+ "absmax_error": 0.008846458047628403,
2090
+ "searched_error": 0.00876232236623764,
2091
+ "improvement": 8.413568139076233e-05,
2092
+ "clipped_row_fraction": 0.3125,
2093
+ "bf16_delta_bytes": 154025912
2094
+ },
2095
+ {
2096
+ "layer": "blocks.34.mlp.fc2",
2097
+ "block": 34,
2098
+ "kind": "mlp.fc2",
2099
+ "shape": [
2100
+ 5376,
2101
+ 14336
2102
+ ],
2103
+ "sample_rows": 64,
2104
+ "absmax_error": 0.009401639923453331,
2105
+ "searched_error": 0.009106581099331379,
2106
+ "improvement": 0.00029505882412195206,
2107
+ "clipped_row_fraction": 0.859375,
2108
+ "bf16_delta_bytes": 77048760
2109
+ },
2110
+ {
2111
+ "layer": "blocks.35.attn.qkv_proj",
2112
+ "block": 35,
2113
+ "kind": "attn.qkv_proj",
2114
+ "shape": [
2115
+ 21504,
2116
+ 5376
2117
+ ],
2118
+ "sample_rows": 64,
2119
+ "absmax_error": 0.00885989610105753,
2120
+ "searched_error": 0.008797014132142067,
2121
+ "improvement": 6.28819689154625e-05,
2122
+ "clipped_row_fraction": 0.21875,
2123
+ "bf16_delta_bytes": 115519416
2124
+ },
2125
+ {
2126
+ "layer": "blocks.35.attn.out_proj",
2127
+ "block": 35,
2128
+ "kind": "attn.out_proj",
2129
+ "shape": [
2130
+ 5376,
2131
+ 7168
2132
+ ],
2133
+ "sample_rows": 64,
2134
+ "absmax_error": 0.009578592143952847,
2135
+ "searched_error": 0.00948287919163704,
2136
+ "improvement": 9.571295231580734e-05,
2137
+ "clipped_row_fraction": 0.328125,
2138
+ "bf16_delta_bytes": 38513592
2139
+ },
2140
+ {
2141
+ "layer": "blocks.35.mlp.fc1",
2142
+ "block": 35,
2143
+ "kind": "mlp.fc1",
2144
+ "shape": [
2145
+ 28672,
2146
+ 5376
2147
+ ],
2148
+ "sample_rows": 64,
2149
+ "absmax_error": 0.008819864131510258,
2150
+ "searched_error": 0.008781999349594116,
2151
+ "improvement": 3.786478191614151e-05,
2152
+ "clipped_row_fraction": 0.140625,
2153
+ "bf16_delta_bytes": 154025912
2154
+ },
2155
+ {
2156
+ "layer": "blocks.35.mlp.fc2",
2157
+ "block": 35,
2158
+ "kind": "mlp.fc2",
2159
+ "shape": [
2160
+ 5376,
2161
+ 14336
2162
+ ],
2163
+ "sample_rows": 64,
2164
+ "absmax_error": 0.009462808258831501,
2165
+ "searched_error": 0.009170095436275005,
2166
+ "improvement": 0.00029271282255649567,
2167
+ "clipped_row_fraction": 0.828125,
2168
+ "bf16_delta_bytes": 77048760
2169
+ },
2170
+ {
2171
+ "layer": "blocks.36.attn.qkv_proj",
2172
+ "block": 36,
2173
+ "kind": "attn.qkv_proj",
2174
+ "shape": [
2175
+ 21504,
2176
+ 5376
2177
+ ],
2178
+ "sample_rows": 64,
2179
+ "absmax_error": 0.008869685232639313,
2180
+ "searched_error": 0.008822268806397915,
2181
+ "improvement": 4.741642624139786e-05,
2182
+ "clipped_row_fraction": 0.171875,
2183
+ "bf16_delta_bytes": 115519416
2184
+ },
2185
+ {
2186
+ "layer": "blocks.36.attn.out_proj",
2187
+ "block": 36,
2188
+ "kind": "attn.out_proj",
2189
+ "shape": [
2190
+ 5376,
2191
+ 7168
2192
+ ],
2193
+ "sample_rows": 64,
2194
+ "absmax_error": 0.010005505755543709,
2195
+ "searched_error": 0.009918516501784325,
2196
+ "improvement": 8.698925375938416e-05,
2197
+ "clipped_row_fraction": 0.28125,
2198
+ "bf16_delta_bytes": 38513592
2199
+ },
2200
+ {
2201
+ "layer": "blocks.36.mlp.fc1",
2202
+ "block": 36,
2203
+ "kind": "mlp.fc1",
2204
+ "shape": [
2205
+ 28672,
2206
+ 5376
2207
+ ],
2208
+ "sample_rows": 64,
2209
+ "absmax_error": 0.008894680067896843,
2210
+ "searched_error": 0.008845840580761433,
2211
+ "improvement": 4.883948713541031e-05,
2212
+ "clipped_row_fraction": 0.203125,
2213
+ "bf16_delta_bytes": 154025912
2214
+ },
2215
+ {
2216
+ "layer": "blocks.36.mlp.fc2",
2217
+ "block": 36,
2218
+ "kind": "mlp.fc2",
2219
+ "shape": [
2220
+ 5376,
2221
+ 14336
2222
+ ],
2223
+ "sample_rows": 64,
2224
+ "absmax_error": 0.00928464811295271,
2225
+ "searched_error": 0.009009410627186298,
2226
+ "improvement": 0.00027523748576641083,
2227
+ "clipped_row_fraction": 0.8125,
2228
+ "bf16_delta_bytes": 77048760
2229
+ },
2230
+ {
2231
+ "layer": "blocks.37.attn.qkv_proj",
2232
+ "block": 37,
2233
+ "kind": "attn.qkv_proj",
2234
+ "shape": [
2235
+ 21504,
2236
+ 5376
2237
+ ],
2238
+ "sample_rows": 64,
2239
+ "absmax_error": 0.008858516812324524,
2240
+ "searched_error": 0.008785505779087543,
2241
+ "improvement": 7.301103323698044e-05,
2242
+ "clipped_row_fraction": 0.25,
2243
+ "bf16_delta_bytes": 115519416
2244
+ },
2245
+ {
2246
+ "layer": "blocks.37.attn.out_proj",
2247
+ "block": 37,
2248
+ "kind": "attn.out_proj",
2249
+ "shape": [
2250
+ 5376,
2251
+ 7168
2252
+ ],
2253
+ "sample_rows": 64,
2254
+ "absmax_error": 0.010057258419692516,
2255
+ "searched_error": 0.009970421902835369,
2256
+ "improvement": 8.683651685714722e-05,
2257
+ "clipped_row_fraction": 0.3125,
2258
+ "bf16_delta_bytes": 38513592
2259
+ },
2260
+ {
2261
+ "layer": "blocks.37.mlp.fc1",
2262
+ "block": 37,
2263
+ "kind": "mlp.fc1",
2264
+ "shape": [
2265
+ 28672,
2266
+ 5376
2267
+ ],
2268
+ "sample_rows": 64,
2269
+ "absmax_error": 0.008871573954820633,
2270
+ "searched_error": 0.008826128207147121,
2271
+ "improvement": 4.5445747673511505e-05,
2272
+ "clipped_row_fraction": 0.171875,
2273
+ "bf16_delta_bytes": 154025912
2274
+ },
2275
+ {
2276
+ "layer": "blocks.37.mlp.fc2",
2277
+ "block": 37,
2278
+ "kind": "mlp.fc2",
2279
+ "shape": [
2280
+ 5376,
2281
+ 14336
2282
+ ],
2283
+ "sample_rows": 64,
2284
+ "absmax_error": 0.009507856331765652,
2285
+ "searched_error": 0.009210675954818726,
2286
+ "improvement": 0.0002971803769469261,
2287
+ "clipped_row_fraction": 0.828125,
2288
+ "bf16_delta_bytes": 77048760
2289
+ },
2290
+ {
2291
+ "layer": "blocks.38.attn.qkv_proj",
2292
+ "block": 38,
2293
+ "kind": "attn.qkv_proj",
2294
+ "shape": [
2295
+ 21504,
2296
+ 5376
2297
+ ],
2298
+ "sample_rows": 64,
2299
+ "absmax_error": 0.008879483677446842,
2300
+ "searched_error": 0.008825055323541164,
2301
+ "improvement": 5.4428353905677795e-05,
2302
+ "clipped_row_fraction": 0.234375,
2303
+ "bf16_delta_bytes": 115519416
2304
+ },
2305
+ {
2306
+ "layer": "blocks.38.attn.out_proj",
2307
+ "block": 38,
2308
+ "kind": "attn.out_proj",
2309
+ "shape": [
2310
+ 5376,
2311
+ 7168
2312
+ ],
2313
+ "sample_rows": 64,
2314
+ "absmax_error": 0.010883932933211327,
2315
+ "searched_error": 0.0107962591573596,
2316
+ "improvement": 8.767377585172653e-05,
2317
+ "clipped_row_fraction": 0.25,
2318
+ "bf16_delta_bytes": 38513592
2319
+ },
2320
+ {
2321
+ "layer": "blocks.38.mlp.fc1",
2322
+ "block": 38,
2323
+ "kind": "mlp.fc1",
2324
+ "shape": [
2325
+ 28672,
2326
+ 5376
2327
+ ],
2328
+ "sample_rows": 64,
2329
+ "absmax_error": 0.00895624328404665,
2330
+ "searched_error": 0.00889838207513094,
2331
+ "improvement": 5.786120891571045e-05,
2332
+ "clipped_row_fraction": 0.21875,
2333
+ "bf16_delta_bytes": 154025912
2334
+ },
2335
+ {
2336
+ "layer": "blocks.38.mlp.fc2",
2337
+ "block": 38,
2338
+ "kind": "mlp.fc2",
2339
+ "shape": [
2340
+ 5376,
2341
+ 14336
2342
+ ],
2343
+ "sample_rows": 64,
2344
+ "absmax_error": 0.009250178001821041,
2345
+ "searched_error": 0.008992554619908333,
2346
+ "improvement": 0.0002576233819127083,
2347
+ "clipped_row_fraction": 0.78125,
2348
+ "bf16_delta_bytes": 77048760
2349
+ },
2350
+ {
2351
+ "layer": "blocks.39.attn.qkv_proj",
2352
+ "block": 39,
2353
+ "kind": "attn.qkv_proj",
2354
+ "shape": [
2355
+ 21504,
2356
+ 5376
2357
+ ],
2358
+ "sample_rows": 64,
2359
+ "absmax_error": 0.008887657895684242,
2360
+ "searched_error": 0.008838411420583725,
2361
+ "improvement": 4.924647510051727e-05,
2362
+ "clipped_row_fraction": 0.21875,
2363
+ "bf16_delta_bytes": 115519416
2364
+ },
2365
+ {
2366
+ "layer": "blocks.39.attn.out_proj",
2367
+ "block": 39,
2368
+ "kind": "attn.out_proj",
2369
+ "shape": [
2370
+ 5376,
2371
+ 7168
2372
+ ],
2373
+ "sample_rows": 64,
2374
+ "absmax_error": 0.009882594458758831,
2375
+ "searched_error": 0.009741849265992641,
2376
+ "improvement": 0.00014074519276618958,
2377
+ "clipped_row_fraction": 0.46875,
2378
+ "bf16_delta_bytes": 38513592
2379
+ },
2380
+ {
2381
+ "layer": "blocks.39.mlp.fc1",
2382
+ "block": 39,
2383
+ "kind": "mlp.fc1",
2384
+ "shape": [
2385
+ 28672,
2386
+ 5376
2387
+ ],
2388
+ "sample_rows": 64,
2389
+ "absmax_error": 0.008798052556812763,
2390
+ "searched_error": 0.008744299411773682,
2391
+ "improvement": 5.3753145039081573e-05,
2392
+ "clipped_row_fraction": 0.21875,
2393
+ "bf16_delta_bytes": 154025912
2394
+ },
2395
+ {
2396
+ "layer": "blocks.39.mlp.fc2",
2397
+ "block": 39,
2398
+ "kind": "mlp.fc2",
2399
+ "shape": [
2400
+ 5376,
2401
+ 14336
2402
+ ],
2403
+ "sample_rows": 64,
2404
+ "absmax_error": 0.009530093520879745,
2405
+ "searched_error": 0.009255670011043549,
2406
+ "improvement": 0.0002744235098361969,
2407
+ "clipped_row_fraction": 0.765625,
2408
+ "bf16_delta_bytes": 77048760
2409
+ },
2410
+ {
2411
+ "layer": "blocks.40.attn.qkv_proj",
2412
+ "block": 40,
2413
+ "kind": "attn.qkv_proj",
2414
+ "shape": [
2415
+ 21504,
2416
+ 5376
2417
+ ],
2418
+ "sample_rows": 64,
2419
+ "absmax_error": 0.008775589056313038,
2420
+ "searched_error": 0.008701545186340809,
2421
+ "improvement": 7.4043869972229e-05,
2422
+ "clipped_row_fraction": 0.234375,
2423
+ "bf16_delta_bytes": 115519416
2424
+ },
2425
+ {
2426
+ "layer": "blocks.40.attn.out_proj",
2427
+ "block": 40,
2428
+ "kind": "attn.out_proj",
2429
+ "shape": [
2430
+ 5376,
2431
+ 7168
2432
+ ],
2433
+ "sample_rows": 64,
2434
+ "absmax_error": 0.009700240567326546,
2435
+ "searched_error": 0.009606442414224148,
2436
+ "improvement": 9.379815310239792e-05,
2437
+ "clipped_row_fraction": 0.296875,
2438
+ "bf16_delta_bytes": 38513592
2439
+ },
2440
+ {
2441
+ "layer": "blocks.40.mlp.fc1",
2442
+ "block": 40,
2443
+ "kind": "mlp.fc1",
2444
+ "shape": [
2445
+ 28672,
2446
+ 5376
2447
+ ],
2448
+ "sample_rows": 64,
2449
+ "absmax_error": 0.008850575424730778,
2450
+ "searched_error": 0.008782174438238144,
2451
+ "improvement": 6.840098649263382e-05,
2452
+ "clipped_row_fraction": 0.234375,
2453
+ "bf16_delta_bytes": 154025912
2454
+ },
2455
+ {
2456
+ "layer": "blocks.40.mlp.fc2",
2457
+ "block": 40,
2458
+ "kind": "mlp.fc2",
2459
+ "shape": [
2460
+ 5376,
2461
+ 14336
2462
+ ],
2463
+ "sample_rows": 64,
2464
+ "absmax_error": 0.00927836075425148,
2465
+ "searched_error": 0.009013586677610874,
2466
+ "improvement": 0.0002647740766406059,
2467
+ "clipped_row_fraction": 0.78125,
2468
+ "bf16_delta_bytes": 77048760
2469
+ },
2470
+ {
2471
+ "layer": "blocks.41.attn.qkv_proj",
2472
+ "block": 41,
2473
+ "kind": "attn.qkv_proj",
2474
+ "shape": [
2475
+ 21504,
2476
+ 5376
2477
+ ],
2478
+ "sample_rows": 64,
2479
+ "absmax_error": 0.008886943571269512,
2480
+ "searched_error": 0.008859841153025627,
2481
+ "improvement": 2.710241824388504e-05,
2482
+ "clipped_row_fraction": 0.125,
2483
+ "bf16_delta_bytes": 115519416
2484
+ },
2485
+ {
2486
+ "layer": "blocks.41.attn.out_proj",
2487
+ "block": 41,
2488
+ "kind": "attn.out_proj",
2489
+ "shape": [
2490
+ 5376,
2491
+ 7168
2492
+ ],
2493
+ "sample_rows": 64,
2494
+ "absmax_error": 0.009859757497906685,
2495
+ "searched_error": 0.009754279628396034,
2496
+ "improvement": 0.00010547786951065063,
2497
+ "clipped_row_fraction": 0.359375,
2498
+ "bf16_delta_bytes": 38513592
2499
+ },
2500
+ {
2501
+ "layer": "blocks.41.mlp.fc1",
2502
+ "block": 41,
2503
+ "kind": "mlp.fc1",
2504
+ "shape": [
2505
+ 28672,
2506
+ 5376
2507
+ ],
2508
+ "sample_rows": 64,
2509
+ "absmax_error": 0.008647102862596512,
2510
+ "searched_error": 0.008588128723204136,
2511
+ "improvement": 5.8974139392375946e-05,
2512
+ "clipped_row_fraction": 0.21875,
2513
+ "bf16_delta_bytes": 154025912
2514
+ },
2515
+ {
2516
+ "layer": "blocks.41.mlp.fc2",
2517
+ "block": 41,
2518
+ "kind": "mlp.fc2",
2519
+ "shape": [
2520
+ 5376,
2521
+ 14336
2522
+ ],
2523
+ "sample_rows": 64,
2524
+ "absmax_error": 0.009476621635258198,
2525
+ "searched_error": 0.009172627702355385,
2526
+ "improvement": 0.00030399393290281296,
2527
+ "clipped_row_fraction": 0.890625,
2528
+ "bf16_delta_bytes": 77048760
2529
+ },
2530
+ {
2531
+ "layer": "blocks.42.attn.qkv_proj",
2532
+ "block": 42,
2533
+ "kind": "attn.qkv_proj",
2534
+ "shape": [
2535
+ 21504,
2536
+ 5376
2537
+ ],
2538
+ "sample_rows": 64,
2539
+ "absmax_error": 0.008771860040724277,
2540
+ "searched_error": 0.00870048813521862,
2541
+ "improvement": 7.13719055056572e-05,
2542
+ "clipped_row_fraction": 0.265625,
2543
+ "bf16_delta_bytes": 115519416
2544
+ },
2545
+ {
2546
+ "layer": "blocks.42.attn.out_proj",
2547
+ "block": 42,
2548
+ "kind": "attn.out_proj",
2549
+ "shape": [
2550
+ 5376,
2551
+ 7168
2552
+ ],
2553
+ "sample_rows": 64,
2554
+ "absmax_error": 0.009836846962571144,
2555
+ "searched_error": 0.009762815199792385,
2556
+ "improvement": 7.4031762778759e-05,
2557
+ "clipped_row_fraction": 0.265625,
2558
+ "bf16_delta_bytes": 38513592
2559
+ },
2560
+ {
2561
+ "layer": "blocks.42.mlp.fc1",
2562
+ "block": 42,
2563
+ "kind": "mlp.fc1",
2564
+ "shape": [
2565
+ 28672,
2566
+ 5376
2567
+ ],
2568
+ "sample_rows": 64,
2569
+ "absmax_error": 0.008824005722999573,
2570
+ "searched_error": 0.008773360401391983,
2571
+ "improvement": 5.064532160758972e-05,
2572
+ "clipped_row_fraction": 0.1875,
2573
+ "bf16_delta_bytes": 154025912
2574
+ },
2575
+ {
2576
+ "layer": "blocks.42.mlp.fc2",
2577
+ "block": 42,
2578
+ "kind": "mlp.fc2",
2579
+ "shape": [
2580
+ 5376,
2581
+ 14336
2582
+ ],
2583
+ "sample_rows": 64,
2584
+ "absmax_error": 0.009384780190885067,
2585
+ "searched_error": 0.009103096090257168,
2586
+ "improvement": 0.00028168410062789917,
2587
+ "clipped_row_fraction": 0.78125,
2588
+ "bf16_delta_bytes": 77048760
2589
+ },
2590
+ {
2591
+ "layer": "blocks.43.attn.qkv_proj",
2592
+ "block": 43,
2593
+ "kind": "attn.qkv_proj",
2594
+ "shape": [
2595
+ 21504,
2596
+ 5376
2597
+ ],
2598
+ "sample_rows": 64,
2599
+ "absmax_error": 0.008804121054708958,
2600
+ "searched_error": 0.008760816417634487,
2601
+ "improvement": 4.330463707447052e-05,
2602
+ "clipped_row_fraction": 0.15625,
2603
+ "bf16_delta_bytes": 115519416
2604
+ },
2605
+ {
2606
+ "layer": "blocks.43.attn.out_proj",
2607
+ "block": 43,
2608
+ "kind": "attn.out_proj",
2609
+ "shape": [
2610
+ 5376,
2611
+ 7168
2612
+ ],
2613
+ "sample_rows": 64,
2614
+ "absmax_error": 0.010491183958947659,
2615
+ "searched_error": 0.010409677401185036,
2616
+ "improvement": 8.150655776262283e-05,
2617
+ "clipped_row_fraction": 0.296875,
2618
+ "bf16_delta_bytes": 38513592
2619
+ },
2620
+ {
2621
+ "layer": "blocks.43.mlp.fc1",
2622
+ "block": 43,
2623
+ "kind": "mlp.fc1",
2624
+ "shape": [
2625
+ 28672,
2626
+ 5376
2627
+ ],
2628
+ "sample_rows": 64,
2629
+ "absmax_error": 0.008889754302799702,
2630
+ "searched_error": 0.00882685650140047,
2631
+ "improvement": 6.289780139923096e-05,
2632
+ "clipped_row_fraction": 0.203125,
2633
+ "bf16_delta_bytes": 154025912
2634
+ },
2635
+ {
2636
+ "layer": "blocks.43.mlp.fc2",
2637
+ "block": 43,
2638
+ "kind": "mlp.fc2",
2639
+ "shape": [
2640
+ 5376,
2641
+ 14336
2642
+ ],
2643
+ "sample_rows": 64,
2644
+ "absmax_error": 0.00917684007436037,
2645
+ "searched_error": 0.008919608779251575,
2646
+ "improvement": 0.00025723129510879517,
2647
+ "clipped_row_fraction": 0.78125,
2648
+ "bf16_delta_bytes": 77048760
2649
+ },
2650
+ {
2651
+ "layer": "blocks.44.attn.qkv_proj",
2652
+ "block": 44,
2653
+ "kind": "attn.qkv_proj",
2654
+ "shape": [
2655
+ 21504,
2656
+ 5376
2657
+ ],
2658
+ "sample_rows": 64,
2659
+ "absmax_error": 0.008927070535719395,
2660
+ "searched_error": 0.008864369243383408,
2661
+ "improvement": 6.270129233598709e-05,
2662
+ "clipped_row_fraction": 0.21875,
2663
+ "bf16_delta_bytes": 115519416
2664
+ },
2665
+ {
2666
+ "layer": "blocks.44.attn.out_proj",
2667
+ "block": 44,
2668
+ "kind": "attn.out_proj",
2669
+ "shape": [
2670
+ 5376,
2671
+ 7168
2672
+ ],
2673
+ "sample_rows": 64,
2674
+ "absmax_error": 0.01023987215012312,
2675
+ "searched_error": 0.010153288953006268,
2676
+ "improvement": 8.65831971168518e-05,
2677
+ "clipped_row_fraction": 0.3125,
2678
+ "bf16_delta_bytes": 38513592
2679
+ },
2680
+ {
2681
+ "layer": "blocks.44.mlp.fc1",
2682
+ "block": 44,
2683
+ "kind": "mlp.fc1",
2684
+ "shape": [
2685
+ 28672,
2686
+ 5376
2687
+ ],
2688
+ "sample_rows": 64,
2689
+ "absmax_error": 0.008789545856416225,
2690
+ "searched_error": 0.008748597465455532,
2691
+ "improvement": 4.094839096069336e-05,
2692
+ "clipped_row_fraction": 0.140625,
2693
+ "bf16_delta_bytes": 154025912
2694
+ },
2695
+ {
2696
+ "layer": "blocks.44.mlp.fc2",
2697
+ "block": 44,
2698
+ "kind": "mlp.fc2",
2699
+ "shape": [
2700
+ 5376,
2701
+ 14336
2702
+ ],
2703
+ "sample_rows": 64,
2704
+ "absmax_error": 0.009537559002637863,
2705
+ "searched_error": 0.009217891842126846,
2706
+ "improvement": 0.00031966716051101685,
2707
+ "clipped_row_fraction": 0.90625,
2708
+ "bf16_delta_bytes": 77048760
2709
+ },
2710
+ {
2711
+ "layer": "blocks.45.attn.qkv_proj",
2712
+ "block": 45,
2713
+ "kind": "attn.qkv_proj",
2714
+ "shape": [
2715
+ 21504,
2716
+ 5376
2717
+ ],
2718
+ "sample_rows": 64,
2719
+ "absmax_error": 0.008663962595164776,
2720
+ "searched_error": 0.008605223149061203,
2721
+ "improvement": 5.8739446103572845e-05,
2722
+ "clipped_row_fraction": 0.21875,
2723
+ "bf16_delta_bytes": 115519416
2724
+ },
2725
+ {
2726
+ "layer": "blocks.45.attn.out_proj",
2727
+ "block": 45,
2728
+ "kind": "attn.out_proj",
2729
+ "shape": [
2730
+ 5376,
2731
+ 7168
2732
+ ],
2733
+ "sample_rows": 64,
2734
+ "absmax_error": 0.010613594204187393,
2735
+ "searched_error": 0.010560794733464718,
2736
+ "improvement": 5.2799470722675323e-05,
2737
+ "clipped_row_fraction": 0.171875,
2738
+ "bf16_delta_bytes": 38513592
2739
+ },
2740
+ {
2741
+ "layer": "blocks.45.mlp.fc1",
2742
+ "block": 45,
2743
+ "kind": "mlp.fc1",
2744
+ "shape": [
2745
+ 28672,
2746
+ 5376
2747
+ ],
2748
+ "sample_rows": 64,
2749
+ "absmax_error": 0.008878356777131557,
2750
+ "searched_error": 0.008829115889966488,
2751
+ "improvement": 4.924088716506958e-05,
2752
+ "clipped_row_fraction": 0.203125,
2753
+ "bf16_delta_bytes": 154025912
2754
+ },
2755
+ {
2756
+ "layer": "blocks.45.mlp.fc2",
2757
+ "block": 45,
2758
+ "kind": "mlp.fc2",
2759
+ "shape": [
2760
+ 5376,
2761
+ 14336
2762
+ ],
2763
+ "sample_rows": 64,
2764
+ "absmax_error": 0.009541809558868408,
2765
+ "searched_error": 0.009230870753526688,
2766
+ "improvement": 0.0003109388053417206,
2767
+ "clipped_row_fraction": 0.890625,
2768
+ "bf16_delta_bytes": 77048760
2769
+ },
2770
+ {
2771
+ "layer": "blocks.46.attn.qkv_proj",
2772
+ "block": 46,
2773
+ "kind": "attn.qkv_proj",
2774
+ "shape": [
2775
+ 21504,
2776
+ 5376
2777
+ ],
2778
+ "sample_rows": 64,
2779
+ "absmax_error": 0.008875416591763496,
2780
+ "searched_error": 0.008823243901133537,
2781
+ "improvement": 5.2172690629959106e-05,
2782
+ "clipped_row_fraction": 0.1875,
2783
+ "bf16_delta_bytes": 115519416
2784
+ },
2785
+ {
2786
+ "layer": "blocks.46.attn.out_proj",
2787
+ "block": 46,
2788
+ "kind": "attn.out_proj",
2789
+ "shape": [
2790
+ 5376,
2791
+ 7168
2792
+ ],
2793
+ "sample_rows": 64,
2794
+ "absmax_error": 0.010151972994208336,
2795
+ "searched_error": 0.010055840946733952,
2796
+ "improvement": 9.613204747438431e-05,
2797
+ "clipped_row_fraction": 0.34375,
2798
+ "bf16_delta_bytes": 38513592
2799
+ },
2800
+ {
2801
+ "layer": "blocks.46.mlp.fc1",
2802
+ "block": 46,
2803
+ "kind": "mlp.fc1",
2804
+ "shape": [
2805
+ 28672,
2806
+ 5376
2807
+ ],
2808
+ "sample_rows": 64,
2809
+ "absmax_error": 0.008902542293071747,
2810
+ "searched_error": 0.008869115263223648,
2811
+ "improvement": 3.3427029848098755e-05,
2812
+ "clipped_row_fraction": 0.125,
2813
+ "bf16_delta_bytes": 154025912
2814
+ },
2815
+ {
2816
+ "layer": "blocks.46.mlp.fc2",
2817
+ "block": 46,
2818
+ "kind": "mlp.fc2",
2819
+ "shape": [
2820
+ 5376,
2821
+ 14336
2822
+ ],
2823
+ "sample_rows": 64,
2824
+ "absmax_error": 0.00948946550488472,
2825
+ "searched_error": 0.009184555150568485,
2826
+ "improvement": 0.0003049103543162346,
2827
+ "clipped_row_fraction": 0.875,
2828
+ "bf16_delta_bytes": 77048760
2829
+ },
2830
+ {
2831
+ "layer": "blocks.47.attn.qkv_proj",
2832
+ "block": 47,
2833
+ "kind": "attn.qkv_proj",
2834
+ "shape": [
2835
+ 21504,
2836
+ 5376
2837
+ ],
2838
+ "sample_rows": 64,
2839
+ "absmax_error": 0.008789182640612125,
2840
+ "searched_error": 0.008733595721423626,
2841
+ "improvement": 5.558691918849945e-05,
2842
+ "clipped_row_fraction": 0.1875,
2843
+ "bf16_delta_bytes": 115519416
2844
+ },
2845
+ {
2846
+ "layer": "blocks.47.attn.out_proj",
2847
+ "block": 47,
2848
+ "kind": "attn.out_proj",
2849
+ "shape": [
2850
+ 5376,
2851
+ 7168
2852
+ ],
2853
+ "sample_rows": 64,
2854
+ "absmax_error": 0.010352781042456627,
2855
+ "searched_error": 0.01026771031320095,
2856
+ "improvement": 8.507072925567627e-05,
2857
+ "clipped_row_fraction": 0.25,
2858
+ "bf16_delta_bytes": 38513592
2859
+ },
2860
+ {
2861
+ "layer": "blocks.47.mlp.fc1",
2862
+ "block": 47,
2863
+ "kind": "mlp.fc1",
2864
+ "shape": [
2865
+ 28672,
2866
+ 5376
2867
+ ],
2868
+ "sample_rows": 64,
2869
+ "absmax_error": 0.008751441724598408,
2870
+ "searched_error": 0.008704815059900284,
2871
+ "improvement": 4.662666469812393e-05,
2872
+ "clipped_row_fraction": 0.21875,
2873
+ "bf16_delta_bytes": 154025912
2874
+ },
2875
+ {
2876
+ "layer": "blocks.47.mlp.fc2",
2877
+ "block": 47,
2878
+ "kind": "mlp.fc2",
2879
+ "shape": [
2880
+ 5376,
2881
+ 14336
2882
+ ],
2883
+ "sample_rows": 64,
2884
+ "absmax_error": 0.00933810044080019,
2885
+ "searched_error": 0.009066533297300339,
2886
+ "improvement": 0.0002715671434998512,
2887
+ "clipped_row_fraction": 0.828125,
2888
+ "bf16_delta_bytes": 77048760
2889
+ },
2890
+ {
2891
+ "layer": "blocks.48.attn.qkv_proj",
2892
+ "block": 48,
2893
+ "kind": "attn.qkv_proj",
2894
+ "shape": [
2895
+ 21504,
2896
+ 5376
2897
+ ],
2898
+ "sample_rows": 64,
2899
+ "absmax_error": 0.00888779480010271,
2900
+ "searched_error": 0.00883910246193409,
2901
+ "improvement": 4.869233816862106e-05,
2902
+ "clipped_row_fraction": 0.203125,
2903
+ "bf16_delta_bytes": 115519416
2904
+ },
2905
+ {
2906
+ "layer": "blocks.48.attn.out_proj",
2907
+ "block": 48,
2908
+ "kind": "attn.out_proj",
2909
+ "shape": [
2910
+ 5376,
2911
+ 7168
2912
+ ],
2913
+ "sample_rows": 64,
2914
+ "absmax_error": 0.009902337566018105,
2915
+ "searched_error": 0.009842359460890293,
2916
+ "improvement": 5.997810512781143e-05,
2917
+ "clipped_row_fraction": 0.234375,
2918
+ "bf16_delta_bytes": 38513592
2919
+ },
2920
+ {
2921
+ "layer": "blocks.48.mlp.fc1",
2922
+ "block": 48,
2923
+ "kind": "mlp.fc1",
2924
+ "shape": [
2925
+ 28672,
2926
+ 5376
2927
+ ],
2928
+ "sample_rows": 64,
2929
+ "absmax_error": 0.008990074507892132,
2930
+ "searched_error": 0.008957264944911003,
2931
+ "improvement": 3.280956298112869e-05,
2932
+ "clipped_row_fraction": 0.125,
2933
+ "bf16_delta_bytes": 154025912
2934
+ },
2935
+ {
2936
+ "layer": "blocks.48.mlp.fc2",
2937
+ "block": 48,
2938
+ "kind": "mlp.fc2",
2939
+ "shape": [
2940
+ 5376,
2941
+ 14336
2942
+ ],
2943
+ "sample_rows": 64,
2944
+ "absmax_error": 0.00943906232714653,
2945
+ "searched_error": 0.00914239976555109,
2946
+ "improvement": 0.0002966625615954399,
2947
+ "clipped_row_fraction": 0.84375,
2948
+ "bf16_delta_bytes": 77048760
2949
+ },
2950
+ {
2951
+ "layer": "blocks.49.attn.qkv_proj",
2952
+ "block": 49,
2953
+ "kind": "attn.qkv_proj",
2954
+ "shape": [
2955
+ 21504,
2956
+ 5376
2957
+ ],
2958
+ "sample_rows": 64,
2959
+ "absmax_error": 0.008817118592560291,
2960
+ "searched_error": 0.00875457189977169,
2961
+ "improvement": 6.254669278860092e-05,
2962
+ "clipped_row_fraction": 0.25,
2963
+ "bf16_delta_bytes": 115519416
2964
+ },
2965
+ {
2966
+ "layer": "blocks.49.attn.out_proj",
2967
+ "block": 49,
2968
+ "kind": "attn.out_proj",
2969
+ "shape": [
2970
+ 5376,
2971
+ 7168
2972
+ ],
2973
+ "sample_rows": 64,
2974
+ "absmax_error": 0.010820906609296799,
2975
+ "searched_error": 0.010764333419501781,
2976
+ "improvement": 5.657318979501724e-05,
2977
+ "clipped_row_fraction": 0.203125,
2978
+ "bf16_delta_bytes": 38513592
2979
+ },
2980
+ {
2981
+ "layer": "blocks.49.mlp.fc1",
2982
+ "block": 49,
2983
+ "kind": "mlp.fc1",
2984
+ "shape": [
2985
+ 28672,
2986
+ 5376
2987
+ ],
2988
+ "sample_rows": 64,
2989
+ "absmax_error": 0.008758306503295898,
2990
+ "searched_error": 0.008717977441847324,
2991
+ "improvement": 4.0329061448574066e-05,
2992
+ "clipped_row_fraction": 0.1875,
2993
+ "bf16_delta_bytes": 154025912
2994
+ },
2995
+ {
2996
+ "layer": "blocks.49.mlp.fc2",
2997
+ "block": 49,
2998
+ "kind": "mlp.fc2",
2999
+ "shape": [
3000
+ 5376,
3001
+ 14336
3002
+ ],
3003
+ "sample_rows": 64,
3004
+ "absmax_error": 0.009797228500247002,
3005
+ "searched_error": 0.009535545483231544,
3006
+ "improvement": 0.00026168301701545715,
3007
+ "clipped_row_fraction": 0.71875,
3008
+ "bf16_delta_bytes": 77048760
3009
+ }
3010
+ ]
3011
+ }
reports/validation_ref2va.json ADDED
@@ -0,0 +1,53 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "release_date": "2026-08-04",
3
+ "checkpoint": "Ref2VA/minimax-h3-ref2va-int8-lean-convrot-table-k16-g4097-quality21.safetensors",
4
+ "bytes": 22484075856,
5
+ "gib": 20.93992741405964,
6
+ "sha256": "f4b2a8b4a53446f5d696a39deb693b09ac7ce37013793a8b19340644e38b154f",
7
+ "tensor_keys": 872,
8
+ "dtype_key_counts": {
9
+ "F32": 282,
10
+ "BF16": 250,
11
+ "I8": 170,
12
+ "U8": 170
13
+ },
14
+ "checks": {
15
+ "safetensors_open": "pass",
16
+ "int8_quant_blobs": "170/170, ConvRot group size 256",
17
+ "bf16_main_policy": "30/30 present; independently selected Ref2VA policy exactly matches FL2VA",
18
+ "nonquant_source_copies": "258/258 tensors, 4146503680 bytes, bitwise equal to raw Ref2VA HF",
19
+ "adaln_table": "FP32 [4097, 16]; independent rebuild max absolute difference 4.547473508864641e-13",
20
+ "midpoint_coordinate_relative_error": 2.1027385628258344e-6,
21
+ "adaln_relative_error_19_timesteps": {
22
+ "block_0": 3.9349262692667253e-7,
23
+ "block_24": 4.4643704200098e-7,
24
+ "block_49": 3.958584784413688e-7,
25
+ "final": 3.423421333081933e-7
26
+ },
27
+ "official_pruned_adaln_relative_error_19_timesteps": {
28
+ "block_0": 1.9679135584738106e-4,
29
+ "block_24": 2.090581983793527e-4,
30
+ "block_49": 2.110520872520283e-4,
31
+ "final": 1.9624050764832646e-4
32
+ },
33
+ "comfyui_cpu_load": "pass on clean unmodified ComfyUI: ModelPatcher -> MiniMaxH3Model; table [4097,16]; no time_embedder",
34
+ "comfyui_tested_commit": "14b05228cef127ce529bc0c08660770d4af3e9a8"
35
+ },
36
+ "sampled_main_matrix_proxy": {
37
+ "method": "deterministic 64 output rows per matrix, all input columns, 200/200 main matrices",
38
+ "official_absmax_mean_relative_error_percent": 0.934510859195143,
39
+ "quality21_effective_mean_relative_error_percent": 0.7639428460970521,
40
+ "relative_reduction_percent": 18.25211675389137,
41
+ "selected_bf16_official_proxy_mean_percent": 1.0683710873126984,
42
+ "remaining_int8_absmax_mean_percent": 0.9108884659979274,
43
+ "remaining_int8_searched_scale_mean_percent": 0.8987562895259437,
44
+ "remaining_int8_relative_reduction_percent": 1.3319058177657648,
45
+ "warning": "This is a weight-reconstruction proxy, not an end-to-end generation or perceptual benchmark."
46
+ },
47
+ "not_yet_tested": [
48
+ "GPU load through ComfyUI",
49
+ "peak VRAM at a target resolution/frame count",
50
+ "end-to-end Ref2VA generation",
51
+ "same-seed visual/audio A/B against BF16 and official pruned checkpoints"
52
+ ]
53
+ }