Defuser 0.0.2__tar.gz → 0.0.3__tar.gz

{defuser-0.0.2 → defuser-0.0.3}/Defuser.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: Defuser
-Version: 0.0.2
+Version: 0.0.3
 Summary: Model defuser helper for HF Transformers.
 Author-email: ModelCloud <qubitium@modelcloud.ai>
 License-Expression: Apache-2.0
@@ -28,7 +28,13 @@ Dynamic: license-file
 <h1 align="center">Defuser</h1>
 </div>
 
-
+<p align="center">
+    <a href="https://github.com/ModelCloud/Defuser/releases" style="text-decoration:none;"><img alt="GitHub release" src="https://img.shields.io/github/release/ModelCloud/Defuser.svg"></a>
+    <a href="https://pypi.org/project/Defuser/" style="text-decoration:none;"><img alt="PyPI - Version" src="https://img.shields.io/pypi/v/Defuser"></a>
+    <a href="https://pepy.tech/projects/Defuser" style="text-decoration:none;"><img src="https://static.pepy.tech/badge/Defuser" alt="PyPI Downloads"></a>
+    <a href="https://github.com/ModelCloud/Defuser/blob/main/LICENSE"><img src="https://img.shields.io/pypi/l/Defuser"></a>
+    <a href="https://huggingface.co/modelcloud/"><img src="https://img.shields.io/badge/🤗%20Hugging%20Face-ModelCloud-%23ff8811.svg"></a>
+</p>
 Model defuser helper for HF Transformers >= 5.0. In HF Transformers 5.x releases, many MoE modules became auto-stacked or auto-fused by new modeling code which has benefits but also downsides. 
 
 * Goal is to provide naive module/layer forwarding code for all models supported by HF transformers where run-time weight and structure level optimizations such weight merging, stacking, fusing are reversed so the model is operating in a simple naive state. 

{defuser-0.0.2 → defuser-0.0.3}/Defuser.egg-info/SOURCES.txt

@@ -10,4 +10,9 @@ Defuser.egg-info/top_level.txt
 defuser/__init__.py
 defuser/defuser.py
 defuser/logger.py
+defuser/modeling/__init__.py
+defuser/modeling/unfused_moe/__init__.py
+defuser/modeling/unfused_moe/qwen3_moe.py
+defuser/utils/__init__.py
+defuser/utils/hf.py
 tests/test_convert_model.py

defuser-0.0.3/Defuser.egg-info/top_level.txt

@@ -0,0 +1 @@
+defuser

{defuser-0.0.2 → defuser-0.0.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: Defuser
-Version: 0.0.2
+Version: 0.0.3
 Summary: Model defuser helper for HF Transformers.
 Author-email: ModelCloud <qubitium@modelcloud.ai>
 License-Expression: Apache-2.0
@@ -28,7 +28,13 @@ Dynamic: license-file
 <h1 align="center">Defuser</h1>
 </div>
 
-
+<p align="center">
+    <a href="https://github.com/ModelCloud/Defuser/releases" style="text-decoration:none;"><img alt="GitHub release" src="https://img.shields.io/github/release/ModelCloud/Defuser.svg"></a>
+    <a href="https://pypi.org/project/Defuser/" style="text-decoration:none;"><img alt="PyPI - Version" src="https://img.shields.io/pypi/v/Defuser"></a>
+    <a href="https://pepy.tech/projects/Defuser" style="text-decoration:none;"><img src="https://static.pepy.tech/badge/Defuser" alt="PyPI Downloads"></a>
+    <a href="https://github.com/ModelCloud/Defuser/blob/main/LICENSE"><img src="https://img.shields.io/pypi/l/Defuser"></a>
+    <a href="https://huggingface.co/modelcloud/"><img src="https://img.shields.io/badge/🤗%20Hugging%20Face-ModelCloud-%23ff8811.svg"></a>
+</p>
 Model defuser helper for HF Transformers >= 5.0. In HF Transformers 5.x releases, many MoE modules became auto-stacked or auto-fused by new modeling code which has benefits but also downsides. 
 
 * Goal is to provide naive module/layer forwarding code for all models supported by HF transformers where run-time weight and structure level optimizations such weight merging, stacking, fusing are reversed so the model is operating in a simple naive state. 

{defuser-0.0.2 → defuser-0.0.3}/README.md

@@ -4,7 +4,13 @@
 <h1 align="center">Defuser</h1>
 </div>
 
-
+<p align="center">
+    <a href="https://github.com/ModelCloud/Defuser/releases" style="text-decoration:none;"><img alt="GitHub release" src="https://img.shields.io/github/release/ModelCloud/Defuser.svg"></a>
+    <a href="https://pypi.org/project/Defuser/" style="text-decoration:none;"><img alt="PyPI - Version" src="https://img.shields.io/pypi/v/Defuser"></a>
+    <a href="https://pepy.tech/projects/Defuser" style="text-decoration:none;"><img src="https://static.pepy.tech/badge/Defuser" alt="PyPI Downloads"></a>
+    <a href="https://github.com/ModelCloud/Defuser/blob/main/LICENSE"><img src="https://img.shields.io/pypi/l/Defuser"></a>
+    <a href="https://huggingface.co/modelcloud/"><img src="https://img.shields.io/badge/🤗%20Hugging%20Face-ModelCloud-%23ff8811.svg"></a>
+</p>
 Model defuser helper for HF Transformers >= 5.0. In HF Transformers 5.x releases, many MoE modules became auto-stacked or auto-fused by new modeling code which has benefits but also downsides. 
 
 * Goal is to provide naive module/layer forwarding code for all models supported by HF transformers where run-time weight and structure level optimizations such weight merging, stacking, fusing are reversed so the model is operating in a simple naive state. 

defuser-0.0.3/defuser/modeling/unfused_moe/qwen3_moe.py

@@ -0,0 +1,67 @@
+# SPDX-FileCopyrightText: 2026 ModelCloud.ai
+# SPDX-FileCopyrightText: 2026 qubitium@modelcloud.ai
+# SPDX-License-Identifier: Apache-2.0
+# Contact: qubitium@modelcloud.ai, x.com/qubitium
+
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+
+
+class LinearQwen3MoeSparseMoeBlock(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        from transformers.models.qwen3_moe.modeling_qwen3_moe import Qwen3MoeMLP
+
+        self.num_experts = config.num_experts
+        self.top_k = config.num_experts_per_tok
+        self.norm_topk_prob = config.norm_topk_prob
+
+        # This must be linear for vllm alignment
+        self.gate = nn.Linear(config.hidden_size, config.num_experts, bias=False)
+        self.experts = nn.ModuleList(
+            [Qwen3MoeMLP(config, intermediate_size=config.moe_intermediate_size) for _ in range(self.num_experts)]
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        """ """
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        # router_logits: (batch * sequence_length, n_experts)
+        router_logits = self.gate(hidden_states)
+
+        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
+        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+        if self.norm_topk_prob:  # only diff with mixtral sparse moe block!
+            routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
+        # we cast back to the input dtype
+        routing_weights = routing_weights.to(hidden_states.dtype)
+
+        final_hidden_states = torch.zeros(
+            (batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device
+        )
+
+        # One hot encode the selected experts to create an expert mask
+        # this will be used to easily index which expert is going to be solicited
+        with torch.no_grad():
+            expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
+
+            # Loop over all available experts in the model and perform the computation on each expert
+            expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero()
+        for expert_idx in expert_hit:
+            if expert_idx == self.num_experts:
+                continue
+            expert_layer = self.experts[expert_idx]
+            idx, top_x = torch.where(expert_mask[expert_idx].squeeze(0))
+
+            # Index the correct hidden states and compute the expert hidden state for
+            # the current expert. We need to make sure to multiply the output hidden
+            # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
+            current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
+            current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None]
+
+            # However `index_add_` only support torch tensors for indexing so we'll use
+            # the `top_x` tensor here.
+            final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
+        final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
+        return final_hidden_states

defuser-0.0.3/defuser/utils/hf.py

@@ -0,0 +1,149 @@
+# SPDX-FileCopyrightText: 2026 ModelCloud.ai
+# SPDX-FileCopyrightText: 2026 qubitium@modelcloud.ai
+# SPDX-License-Identifier: Apache-2.0
+# Contact: qubitium@modelcloud.ai, x.com/qubitium
+
+import torch
+from torch import nn
+
+from packaging import version
+from transformers import AutoConfig
+import transformers
+import os
+import importlib
+
+from typing import Final
+
+from ..logger import logger
+
+MODEL_CONFIG = {
+    "qwen3_moe": {
+        "min_transformers_version": "5.0.0",
+        "checkpoint_mapping": [],
+        "block_patch": [
+            (
+                "transformers.models.qwen3_moe.modeling_qwen3_moe.Qwen3MoeSparseMoeBlock",
+                "defuser.modeling.unfused_moe.qwen3_moe.LinearQwen3MoeSparseMoeBlock",
+            )
+        ],
+    },
+}
+
+_ENV_VAR: Final[str] = "GPTQMODEL_USE_MODELSCOPE"
+
+
+TRUTHFUL = {"1", "true", "yes", "on", "y"}
+
+
+def env_flag(name: str, default: str | bool | None = "0") -> bool:
+    """Return ``True`` when an env var is set to a truthy value."""
+
+    value = os.getenv(name)
+    if value is None:
+        if default is None:
+            return False
+        if isinstance(default, bool):
+            return default
+        value = default
+    return str(value).strip().lower() in TRUTHFUL
+
+def modelscope_requested() -> bool:
+    """
+    Return ``True`` when the user explicitly enabled ModelScope integration
+    via the GPTQMODEL_USE_MODELSCOPE environment variable.
+    """
+    return env_flag(_ENV_VAR, default="0")
+
+
+def get_file_path_via_model_name(model_or_path: str, file_name):
+    from huggingface_hub import hf_hub_download
+
+    # 1) local folder
+    if os.path.isdir(model_or_path):
+        index_path = os.path.join(model_or_path, file_name)
+
+    # 2) HF model name
+    elif not modelscope_requested():
+        index_path = hf_hub_download(
+            repo_id=model_or_path,
+            filename=file_name,
+            repo_type="model",
+        )
+    elif modelscope_requested():
+        from modelscope import snapshot_download  # pylint: disable=E0401
+
+        # ModelSCOPE is different, it returns the folder path
+        folder = snapshot_download(model_or_path, allow_patterns=[file_name])
+        index_path = os.path.join(folder, file_name)
+    else:
+        index_path = None
+
+    return index_path
+
+
+def pre_check_config(model_name: str | torch.nn.Module):
+    if isinstance(model_name, str):
+        config = AutoConfig.from_pretrained(model_name)
+    elif isinstance(model_name, torch.nn.Module):
+        config = getattr(model_name, "config", None)
+        if config is None:
+            return False
+
+    model_type = getattr(config, "model_type", None)
+    if model_type is None or model_type not in MODEL_CONFIG:
+        return False
+
+    cfg = MODEL_CONFIG[model_type]
+
+    min_ver = cfg.get("min_transformers_version")
+    tf_ver = version.parse(transformers.__version__)
+    if min_ver and tf_ver < version.parse(min_ver):
+        return False
+    try:
+        file_path = get_file_path_via_model_name(model_name, "model.safetensors.index.json")
+        if os.path.exists(file_path):
+            import json
+
+            with open(file_path, "r") as f:
+                index_data = json.load(f)
+            model_keys = list(index_data.get("weight_map", {}).keys())
+            for key in model_keys:
+                if "gate_up_proj" in key:
+                    return False
+    except:
+        return True
+    return True
+
+
+def apply_modeling_patch(model: torch.nn.Module) -> bool:
+    res = pre_check_config(model)
+    if not res:
+        return False
+    model_type = getattr(model.config, "model_type")
+    cfg = MODEL_CONFIG[model_type]
+    # patch blocks
+    for orig_path, custom_path in cfg.get("block_patch", []):
+        orig_module_path, orig_class_name = orig_path.rsplit(".", 1)
+        custom_module_path, custom_class_name = custom_path.rsplit(".", 1)
+        try:
+            orig_module = importlib.import_module(orig_module_path)
+            custom_module = importlib.import_module(custom_module_path)
+            custom_class = getattr(custom_module, custom_class_name)
+            orig_class = getattr(orig_module, orig_class_name)
+            names = []
+            for n, m in model.named_modules():
+                if isinstance(m, orig_class):
+                    names.append((n, next(m.parameters()).dtype))
+            for (n, orig_dtype) in names:
+                model.set_submodule(n, custom_class(model.config).to(orig_dtype), True)
+            logger.info(f"Patched module: {orig_path} -> {custom_path}")
+            return True
+        except Exception as e:
+            logger.warning(f"Failed to patch {orig_path}: {e}")
+            return False
+    return False
+
+
+def convert_hf_model(model: nn.Module):
+    apply_modeling_patch(model)
+

{defuser-0.0.2 → defuser-0.0.3}/pyproject.toml

@@ -9,7 +9,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "Defuser"
-version = "0.0.2"
+version = "0.0.3"
 description = "Model defuser helper for HF Transformers."
 readme = "README.md"
 requires-python = ">=3.9"
@@ -38,8 +38,11 @@ dependencies = [
 Homepage = "https://github.com/ModelCloud/Defuser"
 
 [tool.setuptools]
-packages = ["defuser"]
-py-modules = ["setup_utils"]
+include-package-data = false
+
+[tool.setuptools.packages.find]
+where = ["."]
+include = ["defuser*"]
 
 [tool.pytest.ini_options]
 pythonpath = ["."]

{defuser-0.0.2 → defuser-0.0.3}/tests/test_convert_model.py

@@ -3,13 +3,14 @@
 # SPDX-License-Identifier: Apache-2.0
 # Contact: qubitium@modelcloud.ai, x.com/qubitium
 from transformers import AutoModelForCausalLM
+
 from defuser import convert_hf_model
 
 
 def test_qwen3_moe():
     from defuser.modeling.unfused_moe.qwen3_moe import LinearQwen3MoeSparseMoeBlock
 
-    model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-30B-A3B")
+    model = AutoModelForCausalLM.from_pretrained("/monster/data/model/Qwen3-30B-A3B")
 
     assert model.config.model_type == "qwen3_moe"
 

defuser-0.0.2/Defuser.egg-info/top_level.txt

@@ -1,2 +0,0 @@
-defuser
-setup_utils