DmitryDB commited on
Commit
06aba48
·
verified ·
1 Parent(s): adc7972

Remove failed W4 and W8W4 experiments

Browse files

Remove visually broken W4/W8W4 FL2VA and Ref2VA checkpoints and their obsolete validation records. Update the model card and checksums; keep INT8 and NVFP4 releases unchanged.

FL2VA/MiniMax-H3_FL2VA-W4-ConvRot-Offload.safetensors DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:a9d64ad6b7970a9fa7620c8b82cbbec0c8e475310cf6c4c064fcda5d5ee94efb
3
- size 10423705536
 
 
 
 
FL2VA/MiniMax-H3_FL2VA-W4-ConvRot.safetensors DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:73fd34baee103b27f7de5be31b42e8680406816d0eb293f3a9d592fc10947f4a
3
- size 10809057376
 
 
 
 
FL2VA/MiniMax-H3_FL2VA-W8W4-ConvRot.safetensors DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:45aebc6f6ba56a73e32e0cdabb3f039c8346528a9bc4cfc842de6346afdcd658
3
- size 14565748596
 
 
 
 
FL2VA/SHA256SUMS CHANGED
@@ -1,8 +1,5 @@
1
  035312b34614cdc391722dc0d7e9a77a69b2d568ae9079fe629c5efc6457d4a0 MiniMax-H3_FL2VA-INT8-ConvRot-HQ.safetensors
2
  1519c19084ba0387a5f5c3983fcf32aee5923045de1ef4e44c04ca2e2ba486e5 MiniMax-H3_FL2VA-NVFP4.safetensors
3
  1646d1148c0c9a7c2f728c8dbcfe7167e24cc138031feebfd44e04b2daac161f MiniMax-H3_FL2VA-NVFP4-HQ.safetensors
4
- 45aebc6f6ba56a73e32e0cdabb3f039c8346528a9bc4cfc842de6346afdcd658 MiniMax-H3_FL2VA-W8W4-ConvRot.safetensors
5
- 73fd34baee103b27f7de5be31b42e8680406816d0eb293f3a9d592fc10947f4a MiniMax-H3_FL2VA-W4-ConvRot.safetensors
6
  95833a86493b24bcc4d142f541cbc14a7b27de6c5e29a829ff65c40033d91db6 MiniMax-H3_FL2VA-INT8-ConvRot-Lite.safetensors
7
- a9d64ad6b7970a9fa7620c8b82cbbec0c8e475310cf6c4c064fcda5d5ee94efb MiniMax-H3_FL2VA-W4-ConvRot-Offload.safetensors
8
  ad0cf8b50ba92068b5a4477527d5da73412a22db99cded35b07195504e78cf6a MiniMax-H3_FL2VA-INT8-ConvRot.safetensors
 
1
  035312b34614cdc391722dc0d7e9a77a69b2d568ae9079fe629c5efc6457d4a0 MiniMax-H3_FL2VA-INT8-ConvRot-HQ.safetensors
2
  1519c19084ba0387a5f5c3983fcf32aee5923045de1ef4e44c04ca2e2ba486e5 MiniMax-H3_FL2VA-NVFP4.safetensors
3
  1646d1148c0c9a7c2f728c8dbcfe7167e24cc138031feebfd44e04b2daac161f MiniMax-H3_FL2VA-NVFP4-HQ.safetensors
 
 
4
  95833a86493b24bcc4d142f541cbc14a7b27de6c5e29a829ff65c40033d91db6 MiniMax-H3_FL2VA-INT8-ConvRot-Lite.safetensors
 
5
  ad0cf8b50ba92068b5a4477527d5da73412a22db99cded35b07195504e78cf6a MiniMax-H3_FL2VA-INT8-ConvRot.safetensors
README.md CHANGED
@@ -10,7 +10,6 @@ tags:
10
  - comfyui
11
  - quantization
12
  - int8
13
- - w4
14
  - nvfp4
15
  - video
16
  - audio
@@ -49,9 +48,6 @@ Download one FL2VA or Ref2VA checkpoint from the same profile row.
49
  | **INT8 ConvRot HQ** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-INT8-ConvRot-HQ.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-INT8-ConvRot-HQ.safetensors?download=true) | 21.908 GiB | **32 GB+ · RTX 30/40.** 145 INT8 ConvRot + 63 BF16 semantic matrices. Largest BF16 island. A 24 GB RTX 4090 loader test offloaded about 0.955 GiB. |
50
  | **INT8 ConvRot** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-INT8-ConvRot.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-INT8-ConvRot.safetensors?download=true) | 20.940 GiB | **24 GB · RTX 30/40.** 170 INT8 ConvRot + 38 BF16 semantic matrices. Fully resident in the RTX 4090 loader test. |
51
  | **INT8 ConvRot Lite** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-INT8-ConvRot-Lite.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-INT8-ConvRot-Lite.safetensors?download=true) | 20.330 GiB | **24 GB · RTX 30/40.** 185 INT8 ConvRot + 23 BF16 semantic matrices. Leaves more memory for the rest of the workflow. |
52
- | **W8/W4 ConvRot** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-W8W4-ConvRot.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-W8W4-ConvRot.safetensors?download=true) | 13.565 GiB | **16 GB · RTX 30/40.** 86 W8 + 114 W4 main matrices; the eight token-refiner matrices remain BF16. |
53
- | **W4 ConvRot** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-W4-ConvRot.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-W4-ConvRot.safetensors?download=true) | 10.067 GiB | **12 GB · RTX 30/40.** 200 W4 main matrices + 8 INT8 token-refiner matrices. |
54
- | **W4 ConvRot Offload** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-W4-ConvRot-Offload.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-W4-ConvRot-Offload.safetensors?download=true) | 9.708 GiB | **8 GB + CPU offload · RTX 30/40.** All 208 main and token-refiner matrices use W4. |
55
  | **NVFP4 HQ** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-NVFP4-HQ.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-NVFP4-HQ.safetensors?download=true) | 13.597 GiB | **16–24 GB · RTX 50/Blackwell.** 170 NVFP4 + 30 BF16 main matrices; the eight token-refiner matrices remain BF16. |
56
  | **NVFP4** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-NVFP4.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-NVFP4.safetensors?download=true) | 10.862 GiB | **8–12 GB · RTX 50/Blackwell.** All 208 main and token-refiner matrices use block-scaled NVFP4. |
57
 
@@ -76,7 +72,7 @@ VRAM peaks.
76
 
77
  ## Quantization and preserved components
78
 
79
- All 16 diffusion checkpoints:
80
 
81
  - retain all 50 transformer blocks;
82
  - use the fused `qkv_proj = cat(Q,K,V)` layout expected by stock ComfyUI;
@@ -87,7 +83,7 @@ All 16 diffusion checkpoints:
87
  - load without a custom loader or core patch in tested ComfyUI commit
88
  `14b05228`.
89
 
90
- INT8, W8, and W4 weights use ConvRot/Hadamard rotation with group size 256,
91
  per-row FP32 scales, and deterministic scale search.
92
 
93
  | INT8 profile | BF16 attention-output blocks | BF16 MLP `fc2` blocks | Token refiner |
 
10
  - comfyui
11
  - quantization
12
  - int8
 
13
  - nvfp4
14
  - video
15
  - audio
 
48
  | **INT8 ConvRot HQ** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-INT8-ConvRot-HQ.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-INT8-ConvRot-HQ.safetensors?download=true) | 21.908 GiB | **32 GB+ · RTX 30/40.** 145 INT8 ConvRot + 63 BF16 semantic matrices. Largest BF16 island. A 24 GB RTX 4090 loader test offloaded about 0.955 GiB. |
49
  | **INT8 ConvRot** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-INT8-ConvRot.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-INT8-ConvRot.safetensors?download=true) | 20.940 GiB | **24 GB · RTX 30/40.** 170 INT8 ConvRot + 38 BF16 semantic matrices. Fully resident in the RTX 4090 loader test. |
50
  | **INT8 ConvRot Lite** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-INT8-ConvRot-Lite.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-INT8-ConvRot-Lite.safetensors?download=true) | 20.330 GiB | **24 GB · RTX 30/40.** 185 INT8 ConvRot + 23 BF16 semantic matrices. Leaves more memory for the rest of the workflow. |
 
 
 
51
  | **NVFP4 HQ** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-NVFP4-HQ.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-NVFP4-HQ.safetensors?download=true) | 13.597 GiB | **16–24 GB · RTX 50/Blackwell.** 170 NVFP4 + 30 BF16 main matrices; the eight token-refiner matrices remain BF16. |
52
  | **NVFP4** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-NVFP4.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-NVFP4.safetensors?download=true) | 10.862 GiB | **8–12 GB · RTX 50/Blackwell.** All 208 main and token-refiner matrices use block-scaled NVFP4. |
53
 
 
72
 
73
  ## Quantization and preserved components
74
 
75
+ All 10 diffusion checkpoints:
76
 
77
  - retain all 50 transformer blocks;
78
  - use the fused `qkv_proj = cat(Q,K,V)` layout expected by stock ComfyUI;
 
83
  - load without a custom loader or core patch in tested ComfyUI commit
84
  `14b05228`.
85
 
86
+ INT8 weights use ConvRot/Hadamard rotation with group size 256,
87
  per-row FP32 scales, and deterministic scale search.
88
 
89
  | INT8 profile | BF16 attention-output blocks | BF16 MLP `fc2` blocks | Token refiner |
Ref2VA/MiniMax-H3_Ref2VA-W4-ConvRot-Offload.safetensors DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:cc0763a63a1d6f7bb092337712b350fb96507f19be29b1d950de88140183d878
3
- size 10423705536
 
 
 
 
Ref2VA/MiniMax-H3_Ref2VA-W4-ConvRot.safetensors DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:870da69d24a26360c41947c842febbde69ee519428faa1d9a9aad5c4f7d3fa22
3
- size 10809057376
 
 
 
 
Ref2VA/MiniMax-H3_Ref2VA-W8W4-ConvRot.safetensors DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:11a9672f91c8b8a4b18234457f04f5969dcb522ae22a6c7eb8e5d2f472ad496d
3
- size 14565748596
 
 
 
 
Ref2VA/SHA256SUMS CHANGED
@@ -1,8 +1,5 @@
1
  10b4892c0aeb951666c0e8439ebc13d6360d2bfc52457244f37f9b09a3731ee3 MiniMax-H3_Ref2VA-NVFP4-HQ.safetensors
2
- 11a9672f91c8b8a4b18234457f04f5969dcb522ae22a6c7eb8e5d2f472ad496d MiniMax-H3_Ref2VA-W8W4-ConvRot.safetensors
3
  45ec67eaf82718c3f5a23bc3bb77b9d75f1d5a3a2436b6d5a17f183e47048327 MiniMax-H3_Ref2VA-INT8-ConvRot-Lite.safetensors
4
  7693f972cea8b74aa9270550bee4c600cef88ea54f61fcb911ab138fde19b571 MiniMax-H3_Ref2VA-INT8-ConvRot-HQ.safetensors
5
- 870da69d24a26360c41947c842febbde69ee519428faa1d9a9aad5c4f7d3fa22 MiniMax-H3_Ref2VA-W4-ConvRot.safetensors
6
- cc0763a63a1d6f7bb092337712b350fb96507f19be29b1d950de88140183d878 MiniMax-H3_Ref2VA-W4-ConvRot-Offload.safetensors
7
  f4b2a8b4a53446f5d696a39deb693b09ac7ce37013793a8b19340644e38b154f MiniMax-H3_Ref2VA-INT8-ConvRot.safetensors
8
  f970a8a32ded2b4b72f68a624244f8e39d488b1fe0992b9de7133b400dbc95e2 MiniMax-H3_Ref2VA-NVFP4.safetensors
 
1
  10b4892c0aeb951666c0e8439ebc13d6360d2bfc52457244f37f9b09a3731ee3 MiniMax-H3_Ref2VA-NVFP4-HQ.safetensors
 
2
  45ec67eaf82718c3f5a23bc3bb77b9d75f1d5a3a2436b6d5a17f183e47048327 MiniMax-H3_Ref2VA-INT8-ConvRot-Lite.safetensors
3
  7693f972cea8b74aa9270550bee4c600cef88ea54f61fcb911ab138fde19b571 MiniMax-H3_Ref2VA-INT8-ConvRot-HQ.safetensors
 
 
4
  f4b2a8b4a53446f5d696a39deb693b09ac7ce37013793a8b19340644e38b154f MiniMax-H3_Ref2VA-INT8-ConvRot.safetensors
5
  f970a8a32ded2b4b72f68a624244f8e39d488b1fe0992b9de7133b400dbc95e2 MiniMax-H3_Ref2VA-NVFP4.safetensors
reports/release_matrix/loadtest_fl2va_w4-compact12.json DELETED
@@ -1,25 +0,0 @@
1
- {
2
- "checkpoint": "F:\\AI\\hf_publish\\MiniMax-H3-INT8-Lean-ConvRot\\FL2VA\\minimax-h3-fl2va-table-k16-g4097-w4-compact12.safetensors",
3
- "bytes": 10809057376,
4
- "load_device": "cpu",
5
- "stock_comfy_root": "F:\\AI\\comfy_headless\\_stock_comfy_14b",
6
- "elapsed_seconds": 0.4102320671081543,
7
- "checks": {
8
- "model_class": "MiniMaxH3Model",
9
- "table_mode": true,
10
- "has_adaln_t_table": true,
11
- "has_time_embedder": false,
12
- "table_shape": [
13
- 4097,
14
- 16
15
- ],
16
- "has_main_qkv_proj": true,
17
- "has_main_q_proj": false,
18
- "has_refiner_qkv_proj": true,
19
- "quant_formats": {
20
- "int8_tensorwise": 8,
21
- "convrot_w4a4": 200
22
- }
23
- },
24
- "status": "pass"
25
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
reports/release_matrix/loadtest_fl2va_w4-offload8.json DELETED
@@ -1,24 +0,0 @@
1
- {
2
- "checkpoint": "F:\\AI\\hf_publish\\MiniMax-H3-INT8-Lean-ConvRot\\FL2VA\\minimax-h3-fl2va-table-k16-g4097-w4-offload8.safetensors",
3
- "bytes": 10423705536,
4
- "load_device": "cpu",
5
- "stock_comfy_root": "F:\\AI\\comfy_headless\\_stock_comfy_14b",
6
- "elapsed_seconds": 0.44704580307006836,
7
- "checks": {
8
- "model_class": "MiniMaxH3Model",
9
- "table_mode": true,
10
- "has_adaln_t_table": true,
11
- "has_time_embedder": false,
12
- "table_shape": [
13
- 4097,
14
- 16
15
- ],
16
- "has_main_qkv_proj": true,
17
- "has_main_q_proj": false,
18
- "has_refiner_qkv_proj": true,
19
- "quant_formats": {
20
- "convrot_w4a4": 208
21
- }
22
- },
23
- "status": "pass"
24
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
reports/release_matrix/loadtest_fl2va_w8w4-balanced16.json DELETED
@@ -1,25 +0,0 @@
1
- {
2
- "checkpoint": "F:\\AI\\hf_publish\\MiniMax-H3-INT8-Lean-ConvRot\\FL2VA\\minimax-h3-fl2va-table-k16-g4097-w8w4-balanced16.safetensors",
3
- "bytes": 14565748596,
4
- "load_device": "cpu",
5
- "stock_comfy_root": "F:\\AI\\comfy_headless\\_stock_comfy_14b",
6
- "elapsed_seconds": 3.2191989421844482,
7
- "checks": {
8
- "model_class": "MiniMaxH3Model",
9
- "table_mode": true,
10
- "has_adaln_t_table": true,
11
- "has_time_embedder": false,
12
- "table_shape": [
13
- 4097,
14
- 16
15
- ],
16
- "has_main_qkv_proj": true,
17
- "has_main_q_proj": false,
18
- "has_refiner_qkv_proj": true,
19
- "quant_formats": {
20
- "int8_tensorwise": 86,
21
- "convrot_w4a4": 114
22
- }
23
- },
24
- "status": "pass"
25
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
reports/release_matrix/loadtest_ref2va_w4-compact12.json DELETED
@@ -1,25 +0,0 @@
1
- {
2
- "checkpoint": "F:\\AI\\hf_publish\\MiniMax-H3-INT8-Lean-ConvRot\\Ref2VA\\minimax-h3-ref2va-table-k16-g4097-w4-compact12.safetensors",
3
- "bytes": 10809057376,
4
- "load_device": "cpu",
5
- "stock_comfy_root": "F:\\AI\\comfy_headless\\_stock_comfy_14b",
6
- "elapsed_seconds": 0.4829590320587158,
7
- "checks": {
8
- "model_class": "MiniMaxH3Model",
9
- "table_mode": true,
10
- "has_adaln_t_table": true,
11
- "has_time_embedder": false,
12
- "table_shape": [
13
- 4097,
14
- 16
15
- ],
16
- "has_main_qkv_proj": true,
17
- "has_main_q_proj": false,
18
- "has_refiner_qkv_proj": true,
19
- "quant_formats": {
20
- "int8_tensorwise": 8,
21
- "convrot_w4a4": 200
22
- }
23
- },
24
- "status": "pass"
25
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
reports/release_matrix/loadtest_ref2va_w4-offload8.json DELETED
@@ -1,24 +0,0 @@
1
- {
2
- "checkpoint": "F:\\AI\\hf_publish\\MiniMax-H3-INT8-Lean-ConvRot\\Ref2VA\\minimax-h3-ref2va-table-k16-g4097-w4-offload8.safetensors",
3
- "bytes": 10423705536,
4
- "load_device": "cpu",
5
- "stock_comfy_root": "F:\\AI\\comfy_headless\\_stock_comfy_14b",
6
- "elapsed_seconds": 0.4044456481933594,
7
- "checks": {
8
- "model_class": "MiniMaxH3Model",
9
- "table_mode": true,
10
- "has_adaln_t_table": true,
11
- "has_time_embedder": false,
12
- "table_shape": [
13
- 4097,
14
- 16
15
- ],
16
- "has_main_qkv_proj": true,
17
- "has_main_q_proj": false,
18
- "has_refiner_qkv_proj": true,
19
- "quant_formats": {
20
- "convrot_w4a4": 208
21
- }
22
- },
23
- "status": "pass"
24
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
reports/release_matrix/loadtest_ref2va_w8w4-balanced16.json DELETED
@@ -1,25 +0,0 @@
1
- {
2
- "checkpoint": "F:\\AI\\hf_publish\\MiniMax-H3-INT8-Lean-ConvRot\\Ref2VA\\minimax-h3-ref2va-table-k16-g4097-w8w4-balanced16.safetensors",
3
- "bytes": 14565748596,
4
- "load_device": "cpu",
5
- "stock_comfy_root": "F:\\AI\\comfy_headless\\_stock_comfy_14b",
6
- "elapsed_seconds": 3.218238115310669,
7
- "checks": {
8
- "model_class": "MiniMaxH3Model",
9
- "table_mode": true,
10
- "has_adaln_t_table": true,
11
- "has_time_embedder": false,
12
- "table_shape": [
13
- 4097,
14
- 16
15
- ],
16
- "has_main_qkv_proj": true,
17
- "has_main_q_proj": false,
18
- "has_refiner_qkv_proj": true,
19
- "quant_formats": {
20
- "int8_tensorwise": 86,
21
- "convrot_w4a4": 114
22
- }
23
- },
24
- "status": "pass"
25
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
reports/release_matrix/validation_fl2va_w4-compact12.json DELETED
@@ -1,35 +0,0 @@
1
- {
2
- "variant": "fl2va",
3
- "profile": "w4-compact12",
4
- "checkpoint": "F:\\AI\\hf_publish\\MiniMax-H3-INT8-Lean-ConvRot\\FL2VA\\minimax-h3-fl2va-table-k16-g4097-w4-compact12.safetensors",
5
- "bytes": 10809057376,
6
- "gib": 10.066719144582748,
7
- "keys": 948,
8
- "semantic_linear_counts": {
9
- "bf16": 0,
10
- "int8": 8,
11
- "w4": 200,
12
- "nvfp4": 0
13
- },
14
- "sampled_reconstruction": {
15
- "int8": {
16
- "layer": "token_refiner.blocks.0.attn.qkv_proj",
17
- "relative_l2": 0.008800705894827843,
18
- "bitwise": false
19
- },
20
- "w4": {
21
- "layer": "blocks.0.attn.qkv_proj",
22
- "relative_l2": 0.160075381398201,
23
- "bitwise": false
24
- }
25
- },
26
- "adaln_relative_19_timesteps": {
27
- "blocks.0.adaln_proj.linear": 3.5114101137878606e-07,
28
- "blocks.24.adaln_proj.linear": 3.8073906694080506e-07,
29
- "blocks.49.adaln_proj.linear": 3.4442922469679615e-07,
30
- "final_layer.adaln_proj.linear": 2.980012538955634e-07
31
- },
32
- "device": "cpu",
33
- "elapsed_seconds": 111.34770798683167,
34
- "status": "pass"
35
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
reports/release_matrix/validation_fl2va_w4-offload8.json DELETED
@@ -1,30 +0,0 @@
1
- {
2
- "variant": "fl2va",
3
- "profile": "w4-offload8",
4
- "checkpoint": "F:\\AI\\hf_publish\\MiniMax-H3-INT8-Lean-ConvRot\\FL2VA\\minimax-h3-fl2va-table-k16-g4097-w4-offload8.safetensors",
5
- "bytes": 10423705536,
6
- "gib": 9.707832276821136,
7
- "keys": 948,
8
- "semantic_linear_counts": {
9
- "bf16": 0,
10
- "int8": 0,
11
- "w4": 208,
12
- "nvfp4": 0
13
- },
14
- "sampled_reconstruction": {
15
- "w4": {
16
- "layer": "blocks.0.attn.qkv_proj",
17
- "relative_l2": 0.160075381398201,
18
- "bitwise": false
19
- }
20
- },
21
- "adaln_relative_19_timesteps": {
22
- "blocks.0.adaln_proj.linear": 3.5114101137878606e-07,
23
- "blocks.24.adaln_proj.linear": 3.8073906694080506e-07,
24
- "blocks.49.adaln_proj.linear": 3.4442922469679615e-07,
25
- "final_layer.adaln_proj.linear": 2.980012538955634e-07
26
- },
27
- "device": "cpu",
28
- "elapsed_seconds": 94.2013008594513,
29
- "status": "pass"
30
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
reports/release_matrix/validation_fl2va_w8w4-balanced16.json DELETED
@@ -1,40 +0,0 @@
1
- {
2
- "variant": "fl2va",
3
- "profile": "w8w4-balanced16",
4
- "checkpoint": "F:\\AI\\hf_publish\\MiniMax-H3-INT8-Lean-ConvRot\\FL2VA\\minimax-h3-fl2va-table-k16-g4097-w8w4-balanced16.safetensors",
5
- "bytes": 14565748596,
6
- "gib": 13.565410483628511,
7
- "keys": 932,
8
- "semantic_linear_counts": {
9
- "bf16": 8,
10
- "int8": 86,
11
- "w4": 114,
12
- "nvfp4": 0
13
- },
14
- "sampled_reconstruction": {
15
- "bf16": {
16
- "layer": "token_refiner.blocks.0.attn.qkv_proj",
17
- "relative_l2": 0.0,
18
- "bitwise": true
19
- },
20
- "int8": {
21
- "layer": "blocks.0.attn.qkv_proj",
22
- "relative_l2": 0.008814242668449879,
23
- "bitwise": false
24
- },
25
- "w4": {
26
- "layer": "blocks.0.mlp.fc1",
27
- "relative_l2": 0.16060204803943634,
28
- "bitwise": false
29
- }
30
- },
31
- "adaln_relative_19_timesteps": {
32
- "blocks.0.adaln_proj.linear": 3.5114101137878606e-07,
33
- "blocks.24.adaln_proj.linear": 3.8073906694080506e-07,
34
- "blocks.49.adaln_proj.linear": 3.4442922469679615e-07,
35
- "final_layer.adaln_proj.linear": 2.980012538955634e-07
36
- },
37
- "device": "cpu",
38
- "elapsed_seconds": 133.54828667640686,
39
- "status": "pass"
40
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
reports/release_matrix/validation_ref2va_w4-compact12.json DELETED
@@ -1,35 +0,0 @@
1
- {
2
- "variant": "ref2va",
3
- "profile": "w4-compact12",
4
- "checkpoint": "F:\\AI\\hf_publish\\MiniMax-H3-INT8-Lean-ConvRot\\Ref2VA\\minimax-h3-ref2va-table-k16-g4097-w4-compact12.safetensors",
5
- "bytes": 10809057376,
6
- "gib": 10.066719144582748,
7
- "keys": 948,
8
- "semantic_linear_counts": {
9
- "bf16": 0,
10
- "int8": 8,
11
- "w4": 200,
12
- "nvfp4": 0
13
- },
14
- "sampled_reconstruction": {
15
- "int8": {
16
- "layer": "token_refiner.blocks.0.attn.qkv_proj",
17
- "relative_l2": 0.008799571543931961,
18
- "bitwise": false
19
- },
20
- "w4": {
21
- "layer": "blocks.0.attn.qkv_proj",
22
- "relative_l2": 0.1600743532180786,
23
- "bitwise": false
24
- }
25
- },
26
- "adaln_relative_19_timesteps": {
27
- "blocks.0.adaln_proj.linear": 3.457665229689155e-07,
28
- "blocks.24.adaln_proj.linear": 3.6941224834663444e-07,
29
- "blocks.49.adaln_proj.linear": 3.462141933141538e-07,
30
- "final_layer.adaln_proj.linear": 3.0612892487624777e-07
31
- },
32
- "device": "cpu",
33
- "elapsed_seconds": 97.13527059555054,
34
- "status": "pass"
35
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
reports/release_matrix/validation_ref2va_w4-offload8.json DELETED
@@ -1,30 +0,0 @@
1
- {
2
- "variant": "ref2va",
3
- "profile": "w4-offload8",
4
- "checkpoint": "F:\\AI\\hf_publish\\MiniMax-H3-INT8-Lean-ConvRot\\Ref2VA\\minimax-h3-ref2va-table-k16-g4097-w4-offload8.safetensors",
5
- "bytes": 10423705536,
6
- "gib": 9.707832276821136,
7
- "keys": 948,
8
- "semantic_linear_counts": {
9
- "bf16": 0,
10
- "int8": 0,
11
- "w4": 208,
12
- "nvfp4": 0
13
- },
14
- "sampled_reconstruction": {
15
- "w4": {
16
- "layer": "blocks.0.attn.qkv_proj",
17
- "relative_l2": 0.1600743532180786,
18
- "bitwise": false
19
- }
20
- },
21
- "adaln_relative_19_timesteps": {
22
- "blocks.0.adaln_proj.linear": 3.457665229689155e-07,
23
- "blocks.24.adaln_proj.linear": 3.6941224834663444e-07,
24
- "blocks.49.adaln_proj.linear": 3.462141933141538e-07,
25
- "final_layer.adaln_proj.linear": 3.0612892487624777e-07
26
- },
27
- "device": "cpu",
28
- "elapsed_seconds": 55.76402306556702,
29
- "status": "pass"
30
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
reports/release_matrix/validation_ref2va_w8w4-balanced16.json DELETED
@@ -1,40 +0,0 @@
1
- {
2
- "variant": "ref2va",
3
- "profile": "w8w4-balanced16",
4
- "checkpoint": "F:\\AI\\hf_publish\\MiniMax-H3-INT8-Lean-ConvRot\\Ref2VA\\minimax-h3-ref2va-table-k16-g4097-w8w4-balanced16.safetensors",
5
- "bytes": 14565748596,
6
- "gib": 13.565410483628511,
7
- "keys": 932,
8
- "semantic_linear_counts": {
9
- "bf16": 8,
10
- "int8": 86,
11
- "w4": 114,
12
- "nvfp4": 0
13
- },
14
- "sampled_reconstruction": {
15
- "bf16": {
16
- "layer": "token_refiner.blocks.0.attn.qkv_proj",
17
- "relative_l2": 0.0,
18
- "bitwise": true
19
- },
20
- "int8": {
21
- "layer": "blocks.0.attn.qkv_proj",
22
- "relative_l2": 0.008815201930701733,
23
- "bitwise": false
24
- },
25
- "w4": {
26
- "layer": "blocks.0.mlp.fc1",
27
- "relative_l2": 0.16060692071914673,
28
- "bitwise": false
29
- }
30
- },
31
- "adaln_relative_19_timesteps": {
32
- "blocks.0.adaln_proj.linear": 3.457665229689155e-07,
33
- "blocks.24.adaln_proj.linear": 3.6941224834663444e-07,
34
- "blocks.49.adaln_proj.linear": 3.462141933141538e-07,
35
- "final_layer.adaln_proj.linear": 3.0612892487624777e-07
36
- },
37
- "device": "cpu",
38
- "elapsed_seconds": 121.73055958747864,
39
- "status": "pass"
40
- }