Defuser 0.0.21__tar.gz → 0.0.22__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (43) hide show
  1. {defuser-0.0.21 → defuser-0.0.22}/Defuser.egg-info/PKG-INFO +11 -11
  2. {defuser-0.0.21 → defuser-0.0.22}/PKG-INFO +11 -11
  3. {defuser-0.0.21 → defuser-0.0.22}/README.md +10 -10
  4. {defuser-0.0.21 → defuser-0.0.22}/defuser/defuser.py +0 -1
  5. {defuser-0.0.21 → defuser-0.0.22}/defuser/model_registry.py +3 -0
  6. {defuser-0.0.21 → defuser-0.0.22}/defuser/modeling/replace_modules.py +0 -1
  7. {defuser-0.0.21 → defuser-0.0.22}/pyproject.toml +1 -1
  8. {defuser-0.0.21 → defuser-0.0.22}/tests/test_meta_model_defusion.py +11 -0
  9. {defuser-0.0.21 → defuser-0.0.22}/Defuser.egg-info/SOURCES.txt +0 -0
  10. {defuser-0.0.21 → defuser-0.0.22}/Defuser.egg-info/dependency_links.txt +0 -0
  11. {defuser-0.0.21 → defuser-0.0.22}/Defuser.egg-info/requires.txt +0 -0
  12. {defuser-0.0.21 → defuser-0.0.22}/Defuser.egg-info/top_level.txt +0 -0
  13. {defuser-0.0.21 → defuser-0.0.22}/LICENSE +0 -0
  14. {defuser-0.0.21 → defuser-0.0.22}/defuser/__init__.py +0 -0
  15. {defuser-0.0.21 → defuser-0.0.22}/defuser/checkpoint_ops.py +0 -0
  16. {defuser-0.0.21 → defuser-0.0.22}/defuser/modeling/__init__.py +0 -0
  17. {defuser-0.0.21 → defuser-0.0.22}/defuser/modeling/glm4v.py +0 -0
  18. {defuser-0.0.21 → defuser-0.0.22}/defuser/modeling/model_patches.py +0 -0
  19. {defuser-0.0.21 → defuser-0.0.22}/defuser/modeling/moe_experts_interface.py +0 -0
  20. {defuser-0.0.21 → defuser-0.0.22}/defuser/modeling/runtime_defusion.py +0 -0
  21. {defuser-0.0.21 → defuser-0.0.22}/defuser/modeling/unfused_moe/__init__.py +0 -0
  22. {defuser-0.0.21 → defuser-0.0.22}/defuser/modeling/unfused_moe/common.py +0 -0
  23. {defuser-0.0.21 → defuser-0.0.22}/defuser/modeling/unfused_moe/glm4_moe.py +0 -0
  24. {defuser-0.0.21 → defuser-0.0.22}/defuser/modeling/unfused_moe/glm4_moe_lite.py +0 -0
  25. {defuser-0.0.21 → defuser-0.0.22}/defuser/modeling/unfused_moe/mixtral.py +0 -0
  26. {defuser-0.0.21 → defuser-0.0.22}/defuser/modeling/unfused_moe/qwen2_moe.py +0 -0
  27. {defuser-0.0.21 → defuser-0.0.22}/defuser/modeling/unfused_moe/qwen3_moe.py +0 -0
  28. {defuser-0.0.21 → defuser-0.0.22}/defuser/modeling/unfused_moe/qwen3_next.py +0 -0
  29. {defuser-0.0.21 → defuser-0.0.22}/defuser/modeling/unfused_moe/qwen3_omni_moe.py +0 -0
  30. {defuser-0.0.21 → defuser-0.0.22}/defuser/modeling/update_module.py +0 -0
  31. {defuser-0.0.21 → defuser-0.0.22}/defuser/utils/__init__.py +0 -0
  32. {defuser-0.0.21 → defuser-0.0.22}/defuser/utils/common.py +0 -0
  33. {defuser-0.0.21 → defuser-0.0.22}/defuser/utils/device.py +0 -0
  34. {defuser-0.0.21 → defuser-0.0.22}/defuser/utils/hf.py +0 -0
  35. {defuser-0.0.21 → defuser-0.0.22}/defuser/utils/model.py +0 -0
  36. {defuser-0.0.21 → defuser-0.0.22}/setup.cfg +0 -0
  37. {defuser-0.0.21 → defuser-0.0.22}/setup.py +0 -0
  38. {defuser-0.0.21 → defuser-0.0.22}/tests/test_candidate_coverage.py +0 -0
  39. {defuser-0.0.21 → defuser-0.0.22}/tests/test_convert_model.py +0 -0
  40. {defuser-0.0.21 → defuser-0.0.22}/tests/test_device_utils.py +0 -0
  41. {defuser-0.0.21 → defuser-0.0.22}/tests/test_filter_rules.py +0 -0
  42. {defuser-0.0.21 → defuser-0.0.22}/tests/test_moe_experts_interface.py +0 -0
  43. {defuser-0.0.21 → defuser-0.0.22}/tests/test_replace_modules_tracker.py +0 -0
{defuser-0.0.21 → defuser-0.0.22}/Defuser.egg-info/PKG-INFO RENAMED
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: Defuser
3
- Version: 0.0.21
3
+ Version: 0.0.22
4
4
  Summary: Model defuser helper for HF Transformers.
5
5
  Author-email: ModelCloud <qubitium@modelcloud.ai>
6
6
  License-Expression: Apache-2.0
@@ -90,17 +90,17 @@ Defuser currently supports the following `transformers>=5.3.0` `model_type` valu
90
90
 
91
91
  ### 🔄 `convert_model(model)` after load
92
92
 
93
- | Pattern | Supported model types | Defused op performed ⚙️ |
94
- | --- | --- | --- |
93
+ | Pattern | Supported model types | Defused op performed ⚙️ |
94
+ | --- |-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| --- |
95
95
  | Standard routed expert tensors 🧱 | `deepseek_v2`, `dots1`, `ernie4_5_moe`, `ernie4_5_vl_moe`, `exaone_moe`, `flex_olmo`, `glm4_moe_lite`, `glm4v_moe`, `hunyuan_v1_moe`, `jamba`, `laguna`, `lfm2_moe`, `minimax`, `minimax_m2`, `olmoe`, `qwen3_vl_moe`, `solar_open` | Splits fused expert tensors or registered expert buffers into numbered expert `nn.Linear` modules with per-expert `gate_proj`, `up_proj`, and `down_proj`. |
96
- | Mixed sparse and shared experts | `deepseek_v3`, `glm_moe_dsa`, `qwen3_5_moe`, `qwen3_5_moe_text` | Runtime expert tensor defusion for routed experts while preserving the model's shared-expert path. |
97
- | Transposed or packed expert tensors | `gpt_oss`, `phimoe` | Splits transposed fused expert `gate_up_proj` tensors into per-expert `gate_proj` + `up_proj`, preserves expert bias when present, and converts expert tensors into numbered expert `nn.Linear` modules. |
98
- | Flattened expert layout | `dbrx` | Rebuilds the flattened DBRX expert FFN weights into numbered expert `gate_proj`, `up_proj`, and `down_proj` `nn.Linear` modules. |
99
- | Batched expert-input execution | `llama4` | Runtime expert tensor defusion plus preservation of the llama4 batched expert-input execution contract. |
100
- | Non-gated expert MLPs | `nemotron_h` | Converts routed expert tensors into numbered `up_proj` and `down_proj` `nn.Linear` modules for non-gated experts. |
101
- | Parallel expert blocks | `granitemoe`, `granitemoehybrid`, `granitemoeshared`, `jetmoe` | Converts packed expert weight tensors into numbered expert `linear` modules while keeping grouped expert execution intact. |
102
- | Routed experts with identity experts | `longcat_flash` | Defuses routed experts into numbered `gate_proj`, `up_proj`, and `down_proj` modules and preserves zero or identity experts. |
103
- | Fused dense `gate_up_proj` MLPs | `dia`, `glm`, `glm4`, `glm_image`, `glm_ocr`, `phi3`, `phi4_multimodal`, `zamba2` | Splits fused dense `gate_up_proj` layers into `gate_proj` + `up_proj` and updates the block `forward()` to preserve the original MLP math. |
96
+ | Mixed sparse and shared experts | `deepseek_v3`, deepseek_v4`, `glm_moe_dsa`, `qwen3_5_moe`, `qwen3_5_moe_text` | Runtime expert tensor defusion for routed experts while preserving the model's shared-expert path. |
97
+ | Transposed or packed expert tensors | `gpt_oss`, `phimoe` | Splits transposed fused expert `gate_up_proj` tensors into per-expert `gate_proj` + `up_proj`, preserves expert bias when present, and converts expert tensors into numbered expert `nn.Linear` modules. |
98
+ | Flattened expert layout | `dbrx` | Rebuilds the flattened DBRX expert FFN weights into numbered expert `gate_proj`, `up_proj`, and `down_proj` `nn.Linear` modules. |
99
+ | Batched expert-input execution | `llama4` | Runtime expert tensor defusion plus preservation of the llama4 batched expert-input execution contract. |
100
+ | Non-gated expert MLPs | `nemotron_h` | Converts routed expert tensors into numbered `up_proj` and `down_proj` `nn.Linear` modules for non-gated experts. |
101
+ | Parallel expert blocks | `granitemoe`, `granitemoehybrid`, `granitemoeshared`, `jetmoe` | Converts packed expert weight tensors into numbered expert `linear` modules while keeping grouped expert execution intact. |
102
+ | Routed experts with identity experts | `longcat_flash` | Defuses routed experts into numbered `gate_proj`, `up_proj`, and `down_proj` modules and preserves zero or identity experts. |
103
+ | Fused dense `gate_up_proj` MLPs | `dia`, `glm`, `glm4`, `glm_image`, `glm_ocr`, `phi3`, `phi4_multimodal`, `zamba2` | Splits fused dense `gate_up_proj` layers into `gate_proj` + `up_proj` and updates the block `forward()` to preserve the original MLP math. |
104
104
 
105
105
  ## 🔁 Workflow Summary
106
106
 
{defuser-0.0.21 → defuser-0.0.22}/PKG-INFO RENAMED
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: Defuser
3
- Version: 0.0.21
3
+ Version: 0.0.22
4
4
  Summary: Model defuser helper for HF Transformers.
5
5
  Author-email: ModelCloud <qubitium@modelcloud.ai>
6
6
  License-Expression: Apache-2.0
@@ -90,17 +90,17 @@ Defuser currently supports the following `transformers>=5.3.0` `model_type` valu
90
90
 
91
91
  ### 🔄 `convert_model(model)` after load
92
92
 
93
- | Pattern | Supported model types | Defused op performed ⚙️ |
94
- | --- | --- | --- |
93
+ | Pattern | Supported model types | Defused op performed ⚙️ |
94
+ | --- |-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| --- |
95
95
  | Standard routed expert tensors 🧱 | `deepseek_v2`, `dots1`, `ernie4_5_moe`, `ernie4_5_vl_moe`, `exaone_moe`, `flex_olmo`, `glm4_moe_lite`, `glm4v_moe`, `hunyuan_v1_moe`, `jamba`, `laguna`, `lfm2_moe`, `minimax`, `minimax_m2`, `olmoe`, `qwen3_vl_moe`, `solar_open` | Splits fused expert tensors or registered expert buffers into numbered expert `nn.Linear` modules with per-expert `gate_proj`, `up_proj`, and `down_proj`. |
96
- | Mixed sparse and shared experts | `deepseek_v3`, `glm_moe_dsa`, `qwen3_5_moe`, `qwen3_5_moe_text` | Runtime expert tensor defusion for routed experts while preserving the model's shared-expert path. |
97
- | Transposed or packed expert tensors | `gpt_oss`, `phimoe` | Splits transposed fused expert `gate_up_proj` tensors into per-expert `gate_proj` + `up_proj`, preserves expert bias when present, and converts expert tensors into numbered expert `nn.Linear` modules. |
98
- | Flattened expert layout | `dbrx` | Rebuilds the flattened DBRX expert FFN weights into numbered expert `gate_proj`, `up_proj`, and `down_proj` `nn.Linear` modules. |
99
- | Batched expert-input execution | `llama4` | Runtime expert tensor defusion plus preservation of the llama4 batched expert-input execution contract. |
100
- | Non-gated expert MLPs | `nemotron_h` | Converts routed expert tensors into numbered `up_proj` and `down_proj` `nn.Linear` modules for non-gated experts. |
101
- | Parallel expert blocks | `granitemoe`, `granitemoehybrid`, `granitemoeshared`, `jetmoe` | Converts packed expert weight tensors into numbered expert `linear` modules while keeping grouped expert execution intact. |
102
- | Routed experts with identity experts | `longcat_flash` | Defuses routed experts into numbered `gate_proj`, `up_proj`, and `down_proj` modules and preserves zero or identity experts. |
103
- | Fused dense `gate_up_proj` MLPs | `dia`, `glm`, `glm4`, `glm_image`, `glm_ocr`, `phi3`, `phi4_multimodal`, `zamba2` | Splits fused dense `gate_up_proj` layers into `gate_proj` + `up_proj` and updates the block `forward()` to preserve the original MLP math. |
96
+ | Mixed sparse and shared experts | `deepseek_v3`, deepseek_v4`, `glm_moe_dsa`, `qwen3_5_moe`, `qwen3_5_moe_text` | Runtime expert tensor defusion for routed experts while preserving the model's shared-expert path. |
97
+ | Transposed or packed expert tensors | `gpt_oss`, `phimoe` | Splits transposed fused expert `gate_up_proj` tensors into per-expert `gate_proj` + `up_proj`, preserves expert bias when present, and converts expert tensors into numbered expert `nn.Linear` modules. |
98
+ | Flattened expert layout | `dbrx` | Rebuilds the flattened DBRX expert FFN weights into numbered expert `gate_proj`, `up_proj`, and `down_proj` `nn.Linear` modules. |
99
+ | Batched expert-input execution | `llama4` | Runtime expert tensor defusion plus preservation of the llama4 batched expert-input execution contract. |
100
+ | Non-gated expert MLPs | `nemotron_h` | Converts routed expert tensors into numbered `up_proj` and `down_proj` `nn.Linear` modules for non-gated experts. |
101
+ | Parallel expert blocks | `granitemoe`, `granitemoehybrid`, `granitemoeshared`, `jetmoe` | Converts packed expert weight tensors into numbered expert `linear` modules while keeping grouped expert execution intact. |
102
+ | Routed experts with identity experts | `longcat_flash` | Defuses routed experts into numbered `gate_proj`, `up_proj`, and `down_proj` modules and preserves zero or identity experts. |
103
+ | Fused dense `gate_up_proj` MLPs | `dia`, `glm`, `glm4`, `glm_image`, `glm_ocr`, `phi3`, `phi4_multimodal`, `zamba2` | Splits fused dense `gate_up_proj` layers into `gate_proj` + `up_proj` and updates the block `forward()` to preserve the original MLP math. |
104
104
 
105
105
  ## 🔁 Workflow Summary
106
106
 
{defuser-0.0.21 → defuser-0.0.22}/README.md RENAMED
@@ -64,17 +64,17 @@ Defuser currently supports the following `transformers>=5.3.0` `model_type` valu
64
64
 
65
65
  ### 🔄 `convert_model(model)` after load
66
66
 
67
- | Pattern | Supported model types | Defused op performed ⚙️ |
68
- | --- | --- | --- |
67
+ | Pattern | Supported model types | Defused op performed ⚙️ |
68
+ | --- |-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| --- |
69
69
  | Standard routed expert tensors 🧱 | `deepseek_v2`, `dots1`, `ernie4_5_moe`, `ernie4_5_vl_moe`, `exaone_moe`, `flex_olmo`, `glm4_moe_lite`, `glm4v_moe`, `hunyuan_v1_moe`, `jamba`, `laguna`, `lfm2_moe`, `minimax`, `minimax_m2`, `olmoe`, `qwen3_vl_moe`, `solar_open` | Splits fused expert tensors or registered expert buffers into numbered expert `nn.Linear` modules with per-expert `gate_proj`, `up_proj`, and `down_proj`. |
70
- | Mixed sparse and shared experts | `deepseek_v3`, `glm_moe_dsa`, `qwen3_5_moe`, `qwen3_5_moe_text` | Runtime expert tensor defusion for routed experts while preserving the model's shared-expert path. |
71
- | Transposed or packed expert tensors | `gpt_oss`, `phimoe` | Splits transposed fused expert `gate_up_proj` tensors into per-expert `gate_proj` + `up_proj`, preserves expert bias when present, and converts expert tensors into numbered expert `nn.Linear` modules. |
72
- | Flattened expert layout | `dbrx` | Rebuilds the flattened DBRX expert FFN weights into numbered expert `gate_proj`, `up_proj`, and `down_proj` `nn.Linear` modules. |
73
- | Batched expert-input execution | `llama4` | Runtime expert tensor defusion plus preservation of the llama4 batched expert-input execution contract. |
74
- | Non-gated expert MLPs | `nemotron_h` | Converts routed expert tensors into numbered `up_proj` and `down_proj` `nn.Linear` modules for non-gated experts. |
75
- | Parallel expert blocks | `granitemoe`, `granitemoehybrid`, `granitemoeshared`, `jetmoe` | Converts packed expert weight tensors into numbered expert `linear` modules while keeping grouped expert execution intact. |
76
- | Routed experts with identity experts | `longcat_flash` | Defuses routed experts into numbered `gate_proj`, `up_proj`, and `down_proj` modules and preserves zero or identity experts. |
77
- | Fused dense `gate_up_proj` MLPs | `dia`, `glm`, `glm4`, `glm_image`, `glm_ocr`, `phi3`, `phi4_multimodal`, `zamba2` | Splits fused dense `gate_up_proj` layers into `gate_proj` + `up_proj` and updates the block `forward()` to preserve the original MLP math. |
70
+ | Mixed sparse and shared experts | `deepseek_v3`, deepseek_v4`, `glm_moe_dsa`, `qwen3_5_moe`, `qwen3_5_moe_text` | Runtime expert tensor defusion for routed experts while preserving the model's shared-expert path. |
71
+ | Transposed or packed expert tensors | `gpt_oss`, `phimoe` | Splits transposed fused expert `gate_up_proj` tensors into per-expert `gate_proj` + `up_proj`, preserves expert bias when present, and converts expert tensors into numbered expert `nn.Linear` modules. |
72
+ | Flattened expert layout | `dbrx` | Rebuilds the flattened DBRX expert FFN weights into numbered expert `gate_proj`, `up_proj`, and `down_proj` `nn.Linear` modules. |
73
+ | Batched expert-input execution | `llama4` | Runtime expert tensor defusion plus preservation of the llama4 batched expert-input execution contract. |
74
+ | Non-gated expert MLPs | `nemotron_h` | Converts routed expert tensors into numbered `up_proj` and `down_proj` `nn.Linear` modules for non-gated experts. |
75
+ | Parallel expert blocks | `granitemoe`, `granitemoehybrid`, `granitemoeshared`, `jetmoe` | Converts packed expert weight tensors into numbered expert `linear` modules while keeping grouped expert execution intact. |
76
+ | Routed experts with identity experts | `longcat_flash` | Defuses routed experts into numbered `gate_proj`, `up_proj`, and `down_proj` modules and preserves zero or identity experts. |
77
+ | Fused dense `gate_up_proj` MLPs | `dia`, `glm`, `glm4`, `glm_image`, `glm_ocr`, `phi3`, `phi4_multimodal`, `zamba2` | Splits fused dense `gate_up_proj` layers into `gate_proj` + `up_proj` and updates the block `forward()` to preserve the original MLP math. |
78
78
 
79
79
  ## 🔁 Workflow Summary
80
80
 
{defuser-0.0.21 → defuser-0.0.22}/defuser/defuser.py RENAMED
@@ -75,7 +75,6 @@ def replace_fused_blocks(model_type: str) -> bool:
75
75
  try:
76
76
  orig_module = importlib.import_module(orig_module_path)
77
77
  custom_module = importlib.import_module(custom_module_path)
78
- print("orig_module", orig_module, orig_class_name)
79
78
  # Validate class existence before patching
80
79
  if not hasattr(orig_module, orig_class_name):
81
80
  raise PatchError(f"Original class[{orig_class_name}] not found: {orig_module}")
{defuser-0.0.21 → defuser-0.0.22}/defuser/model_registry.py RENAMED
@@ -25,6 +25,9 @@ MODEL_CONFIG = {
25
25
  "deepseek_v3": {
26
26
  "min_transformers_version": MIN_SUPPORTED_TRANSFORMERS_VERSION,
27
27
  },
28
+ "deepseek_v4": {
29
+ "min_transformers_version": MIN_SUPPORTED_TRANSFORMERS_VERSION,
30
+ },
28
31
  "dia": {
29
32
  "min_transformers_version": MIN_SUPPORTED_TRANSFORMERS_VERSION,
30
33
  },
{defuser-0.0.21 → defuser-0.0.22}/defuser/modeling/replace_modules.py RENAMED
@@ -325,7 +325,6 @@ def _apply_custom_replacements(
325
325
  module,
326
326
  model.config,
327
327
  ).to(orig_dtype)
328
- print("replacement", replacement)
329
328
  model.set_submodule(name, replacement)
330
329
  replaced.append((name, replacement_cls))
331
330
  else:
{defuser-0.0.21 → defuser-0.0.22}/pyproject.toml RENAMED
@@ -9,7 +9,7 @@ build-backend = "setuptools.build_meta"
9
9
 
10
10
  [project]
11
11
  name = "Defuser"
12
- version = "0.0.21"
12
+ version = "0.0.22"
13
13
  description = "Model defuser helper for HF Transformers."
14
14
  readme = "README.md"
15
15
  requires-python = ">=3.9"
{defuser-0.0.21 → defuser-0.0.22}/tests/test_meta_model_defusion.py RENAMED
@@ -331,6 +331,17 @@ META_MODEL_CASES = [
331
331
  "validator": "experts",
332
332
  "min_targets": 2,
333
333
  },
334
+ {
335
+ "model_type": "deepseek_v4",
336
+ "mode": "convert",
337
+ "model_module": "transformers.models.deepseek_v4.modeling_deepseek_v4",
338
+ "model_class": "DeepseekV4ForCausalLM",
339
+ "config_module": "transformers.models.deepseek_v4.configuration_deepseek_v4",
340
+ "config_class": "DeepseekV4Config",
341
+ "target_class_paths": ("transformers.models.deepseek_v4.modeling_deepseek_v4.DeepseekV4Experts",),
342
+ "validator": "experts",
343
+ "min_targets": 2,
344
+ },
334
345
  {
335
346
  "model_type": "dia",
336
347
  "mode": "convert",
{defuser-0.0.21 → defuser-0.0.22}/LICENSE RENAMED
File without changes
{defuser-0.0.21 → defuser-0.0.22}/defuser/__init__.py RENAMED
File without changes
{defuser-0.0.21 → defuser-0.0.22}/defuser/utils/hf.py RENAMED
File without changes
{defuser-0.0.21 → defuser-0.0.22}/setup.cfg RENAMED
File without changes
{defuser-0.0.21 → defuser-0.0.22}/setup.py RENAMED
File without changes