transformers 5.6.0__tar.gz → 5.6.2__tar.gz

{transformers-5.6.0/src/transformers.egg-info → transformers-5.6.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: transformers
-Version: 5.6.0
+Version: 5.6.2
 Summary: Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference and training.
 Home-page: https://github.com/huggingface/transformers
 Author: The Hugging Face team (past and future) with the help of all our contributors (https://github.com/huggingface/transformers/graphs/contributors)

{transformers-5.6.0 → transformers-5.6.2}/setup.py

@@ -328,7 +328,7 @@ if __name__ == "__main__":
 
     setup(
         name="transformers",
-        version="5.6.0",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
+        version="5.6.2",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
         author="The Hugging Face team (past and future) with the help of all our contributors (https://github.com/huggingface/transformers/graphs/contributors)",
         author_email="transformers@huggingface.co",
         description="Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference and training.",

{transformers-5.6.0 → transformers-5.6.2}/src/transformers/__init__.py

@@ -18,7 +18,7 @@
 # to defer the actual importing for when the objects are requested. This way `import transformers` provides the names
 # in the namespace without actually importing anything (and especially none of the backends).
 
-__version__ = "5.6.0"
+__version__ = "5.6.2"
 
 import importlib
 import sys

{transformers-5.6.0 → transformers-5.6.2}/src/transformers/conversion_mapping.py

@@ -75,7 +75,6 @@ _MODEL_TO_CONVERSION_PATTERN = {
     "qwen2_5_vl": "qwen2_vl",
     "sam3_tracker_video": "sam3_tracker",
     "pp_chart2table": "llava",
-    "qwen3_5_moe_text": "qwen3_5_text",
     "altclip_vision_model": "clip_vision_model",
     "chinese_clip_vision_model": "clip_vision_model",
     "clipseg_vision_model": "clip_vision_model",
@@ -600,6 +599,8 @@ def _build_checkpoint_conversion_mapping():
         WeightRenaming(source_patterns=r"mlp\.expert_bias", target_patterns="mlp.e_score_correction_bias"),
         WeightRenaming(source_patterns=r"mlp\.shared_mlp\.", target_patterns="mlp.shared_experts."),
     ]
+    mapping["qwen3_5_moe_text"] = mapping["qwen3_5_text"].copy()
+    mapping["qwen3_5_moe_text"] += mapping["qwen2_moe"].copy()
 
     for model_type, base_pattern in _MODEL_TO_CONVERSION_PATTERN.items():
         if model_type in mapping:

{transformers-5.6.0 → transformers-5.6.2}/src/transformers/integrations/finegrained_fp8.py

@@ -53,6 +53,13 @@ deepgemm_per_token_cast_to_fp8 = None
 _deepgemm_available = None
 
 
+def _first_attr(obj, *names):
+    for name in names:
+        if hasattr(obj, name):
+            return getattr(obj, name)
+    raise AttributeError(f"{type(obj).__name__} has none of: {names}")
+
+
 def _load_triton_kernel():
     """Lazily load the finegrained-fp8 Triton kernel and extract functions.
 
@@ -74,10 +81,10 @@ def _load_triton_kernel():
     _triton_available = False  # mark attempted before any early exit
 
     kernel = lazy_load_kernel("finegrained-fp8")
-    triton_fp8_matmul = getattr(kernel, "w8a8_fp8_matmul")
-    triton_fp8_act_quant = getattr(kernel, "fp8_act_quant")
-    triton_batched_fp8_matmul = getattr(kernel, "w8a8_fp8_matmul_batched")
-    triton_grouped_fp8_matmul = getattr(kernel, "w8a8_fp8_matmul_grouped")
+    triton_fp8_matmul = getattr(kernel, "w8a8_fp8_matmul", None)
+    triton_fp8_act_quant = getattr(kernel, "fp8_act_quant", None)
+    triton_batched_fp8_matmul = getattr(kernel, "w8a8_fp8_matmul_batched", None)
+    triton_grouped_fp8_matmul = getattr(kernel, "w8a8_fp8_matmul_grouped", None)
 
     missing = [
         name
@@ -137,8 +144,8 @@ def _load_deepgemm_kernel():
         )
 
     kernel = lazy_load_kernel("deep-gemm")
-    deepgemm_fp8_matmul = getattr(kernel, "fp8_gemm_nt")
-    deepgemm_grouped_fp8_matmul = getattr(kernel, "m_grouped_fp8_gemm_nt_contiguous")
+    deepgemm_fp8_matmul = getattr(kernel, "fp8_gemm_nt", None)
+    deepgemm_grouped_fp8_matmul = getattr(kernel, "m_grouped_fp8_gemm_nt_contiguous", None)
     deepgemm_per_token_cast_to_fp8 = resolve_internal_import(kernel, chained_path="utils.per_token_cast_to_fp8")
 
     missing = [
@@ -596,9 +603,9 @@ class FP8Experts(nn.Module):
         self.block_size = block_size
         self.hidden_dim = config.hidden_size
         self.activation_scheme = activation_scheme
-        self.num_experts = getattr(config, "num_local_experts", config.num_experts)
-        self.intermediate_dim = getattr(config, "moe_intermediate_size", config.intermediate_size)
-        self.act_fn = ACT2FN[getattr(config, "hidden_activation", config.hidden_act)]
+        self.num_experts = _first_attr(config, "num_local_experts", "num_experts")
+        self.intermediate_dim = _first_attr(config, "moe_intermediate_size", "intermediate_size")
+        self.act_fn = ACT2FN[_first_attr(config, "hidden_activation", "hidden_act")]
 
         if self.has_gate:
             gu_proj_out, gu_proj_in = 2 * self.intermediate_dim, self.hidden_dim

{transformers-5.6.0 → transformers-5.6.2}/src/transformers/integrations/flash_attention.py

@@ -81,7 +81,11 @@ def flash_attention_forward(
         target_dtype=target_dtype,
         attn_implementation=module.config._attn_implementation,
         layer_idx=module.layer_idx if hasattr(module, "layer_idx") else None,
-        s_aux=s_aux.to(query.dtype),  # FA only accepts half precision
+        s_aux=(
+            s_aux.to(query.dtype)  # FA only accepts half precision
+            if s_aux is not None
+            else None
+        ),
         **kwargs,
     )
 

{transformers-5.6.0 → transformers-5.6.2/src/transformers.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: transformers
-Version: 5.6.0
+Version: 5.6.2
 Summary: Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference and training.
 Home-page: https://github.com/huggingface/transformers
 Author: The Hugging Face team (past and future) with the help of all our contributors (https://github.com/huggingface/transformers/graphs/contributors)