transformers 5.0.0rc0__py3-none-any.whl → 5.0.0rc2__py3-none-any.whl

transformers/models/qwen2_vl/modeling_qwen2_vl.py

@@ -28,6 +28,7 @@ import torch.nn as nn
 import torch.nn.functional as F
 from torch.nn import LayerNorm
 
+from ... import initialization as init
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...generation import GenerationMixin
@@ -42,9 +43,9 @@ from ...utils import (
     TransformersKwargs,
     auto_docstring,
     can_return_tuple,
-    is_torchdynamo_compiling,
     logging,
 )
+from ...utils.generic import maybe_autocast
 from ..qwen2.modeling_qwen2 import (
     Qwen2RMSNorm,
 )
@@ -125,7 +126,7 @@ class Qwen2VLRotaryEmbedding(nn.Module):
         inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
 
         self.register_buffer("inv_freq", inv_freq, persistent=False)
-        self.original_inv_freq = inv_freq
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
 
     @staticmethod
     def compute_default_rope_parameters(
@@ -165,7 +166,7 @@ class Qwen2VLRotaryEmbedding(nn.Module):
         position_ids_expanded = position_ids[:, :, None, :].float()  # shape (3, bs, 1, positions)
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
             freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(2, 3)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
@@ -246,6 +247,8 @@ class VisionRotaryEmbedding(nn.Module):
 
     def __init__(self, dim: int, theta: float = 10000.0) -> None:
         super().__init__()
+        self.dim = dim
+        self.theta = theta
         inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
         self.register_buffer("inv_freq", inv_freq, persistent=False)
 
@@ -384,8 +387,8 @@ class VisionAttention(nn.Module):
         if self.config._attn_implementation != "eager":
             attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
 
-        if self.config._attn_implementation == "flash_attention_2":
-            # Flash Attention 2: Use cu_seqlens for variable length attention
+        if "flash" in self.config._attn_implementation:
+            # Flash Attention: Use cu_seqlens for variable length attention
             max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
             attn_output, _ = attention_interface(
                 self,
@@ -665,6 +668,12 @@ class Qwen2VLPreTrainedModel(PreTrainedModel):
     _can_compile_fullgraph = True
     _supports_attention_backend = True
 
+    def _init_weights(self, module):
+        super()._init_weights(module)
+        if isinstance(module, VisionRotaryEmbedding):
+            inv_freq = 1.0 / (module.theta ** (torch.arange(0, module.dim, 2, dtype=torch.float) / module.dim))
+            init.copy_(module.inv_freq, inv_freq)
+
 
 @auto_docstring
 class Qwen2VisionTransformerPretrainedModel(Qwen2VLPreTrainedModel):
@@ -693,6 +702,8 @@ class Qwen2VisionTransformerPretrainedModel(Qwen2VLPreTrainedModel):
         )
         self.gradient_checkpointing = False
 
+        self.post_init()
+
     def get_dtype(self) -> torch.dtype:
         return self.blocks[0].mlp.fc2.weight.dtype
 
@@ -1222,7 +1233,8 @@ class Qwen2VLModel(Qwen2VLPreTrainedModel):
             inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)
 
         if position_ids is None:
-            if self.rope_deltas is None or cache_position is None or cache_position[0] == 0:
+            past_key_values_length = 0 if past_key_values is None else past_key_values.get_seq_length()
+            if self.rope_deltas is None or past_key_values_length == 0:
                 position_ids, rope_deltas = self.get_rope_index(
                     input_ids, image_grid_thw, video_grid_thw, attention_mask
                 )
@@ -1232,10 +1244,7 @@ class Qwen2VLModel(Qwen2VLPreTrainedModel):
                 batch_size, seq_length, _ = inputs_embeds.shape
                 position_ids = torch.arange(seq_length, device=inputs_embeds.device)
                 position_ids = position_ids.view(1, 1, -1).expand(3, batch_size, -1)
-                if cache_position is not None:
-                    delta = (cache_position[0] + self.rope_deltas).to(inputs_embeds.device)
-                else:
-                    delta = torch.zeros((batch_size, seq_length), device=inputs_embeds.device)
+                delta = (past_key_values_length + self.rope_deltas).to(inputs_embeds.device)
                 delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)
                 position_ids = position_ids + delta.to(position_ids.device)
 
@@ -1418,6 +1427,7 @@ class Qwen2VLForConditionalGeneration(Qwen2VLPreTrainedModel, GenerationMixin):
         pixel_values_videos=None,
         image_grid_thw=None,
         video_grid_thw=None,
+        is_first_iteration=False,
         **kwargs,
     ):
         # Overwritten -- in specific circumstances we don't want to forward image inputs to the model
@@ -1434,6 +1444,7 @@ class Qwen2VLForConditionalGeneration(Qwen2VLPreTrainedModel, GenerationMixin):
             image_grid_thw=image_grid_thw,
             video_grid_thw=video_grid_thw,
             use_cache=use_cache,
+            is_first_iteration=is_first_iteration,
             **kwargs,
         )
 
@@ -1443,15 +1454,7 @@ class Qwen2VLForConditionalGeneration(Qwen2VLPreTrainedModel, GenerationMixin):
             # When compiling, we can't check tensor values thus we check only input length
             # It is safe to assume that `length!=1` means we're in pre-fill because compiled
             # models currently cannot do asssisted decoding
-            prefill_compiled_stage = is_torchdynamo_compiling() and (
-                (input_ids is not None and input_ids.shape[1] != 1)
-                or (inputs_embeds is not None and inputs_embeds.shape[1] != 1)
-            )
-            prefill_noncompiled_stage = not is_torchdynamo_compiling() and (
-                (cache_position is not None and cache_position[0] == 0)
-                or (past_key_values is None or past_key_values.get_seq_length() == 0)
-            )
-            if (prefill_compiled_stage or prefill_noncompiled_stage) or self.model.rope_deltas is None:
+            if model_inputs["cache_position"][0] == 0 or self.model.rope_deltas is None:
                 vision_positions, rope_deltas = self.model.get_rope_index(
                     model_inputs.get("input_ids", None),
                     image_grid_thw=image_grid_thw,
@@ -1473,7 +1476,7 @@ class Qwen2VLForConditionalGeneration(Qwen2VLPreTrainedModel, GenerationMixin):
             text_positions = model_inputs["position_ids"][None, ...]
             model_inputs["position_ids"] = torch.cat([text_positions, vision_positions], dim=0)
 
-        if model_inputs["cache_position"][0] != 0:
+        if not is_first_iteration and use_cache:
             model_inputs["pixel_values"] = None
             model_inputs["pixel_values_videos"] = None
 

transformers/models/qwen3/modeling_qwen3.py

@@ -28,7 +28,7 @@ from torch import nn
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...generation import GenerationMixin
-from ...integrations import use_kernel_forward_from_hub, use_kernel_func_from_hub
+from ...integrations import use_kernel_forward_from_hub, use_kernel_func_from_hub, use_kernelized_func
 from ...masking_utils import create_causal_mask, create_sliding_window_causal_mask
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_layers import (
@@ -42,7 +42,7 @@ from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
 from ...utils import TransformersKwargs, auto_docstring, can_return_tuple
-from ...utils.generic import check_model_inputs
+from ...utils.generic import check_model_inputs, maybe_autocast
 from .configuration_qwen3 import Qwen3Config
 
 
@@ -100,7 +100,7 @@ class Qwen3RotaryEmbedding(nn.Module):
         inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
 
         self.register_buffer("inv_freq", inv_freq, persistent=False)
-        self.original_inv_freq = inv_freq
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
 
     @staticmethod
     def compute_default_rope_parameters(
@@ -139,7 +139,7 @@ class Qwen3RotaryEmbedding(nn.Module):
         position_ids_expanded = position_ids[:, None, :].float()
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
             freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
@@ -221,6 +221,7 @@ def eager_attention_forward(
     return attn_output, attn_weights
 
 
+@use_kernelized_func(apply_rotary_pos_emb)
 class Qwen3Attention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
@@ -247,7 +248,6 @@ class Qwen3Attention(nn.Module):
         self.o_proj = nn.Linear(
             config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
         )
-        self.rotary_fn = apply_rotary_pos_emb
         self.q_norm = Qwen3RMSNorm(self.head_dim, eps=config.rms_norm_eps)  # unlike olmo, only on the head dim!
         self.k_norm = Qwen3RMSNorm(self.head_dim, eps=config.rms_norm_eps)  # thus post q_norm does not need reshape
         self.sliding_window = config.sliding_window if self.layer_type == "sliding_attention" else None

transformers/models/qwen3_moe/modeling_qwen3_moe.py

@@ -30,7 +30,12 @@ from ... import initialization as init
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...generation import GenerationMixin
-from ...integrations import use_kernel_forward_from_hub, use_kernel_func_from_hub
+from ...integrations import (
+    use_experts_implementation,
+    use_kernel_forward_from_hub,
+    use_kernel_func_from_hub,
+    use_kernelized_func,
+)
 from ...masking_utils import create_causal_mask, create_sliding_window_causal_mask
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_layers import (
@@ -43,8 +48,8 @@ from ...modeling_outputs import MoeCausalLMOutputWithPast, MoeModelOutputWithPas
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
-from ...utils import TransformersKwargs, auto_docstring, can_return_tuple
-from ...utils.generic import OutputRecorder, check_model_inputs
+from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, is_grouped_mm_available
+from ...utils.generic import OutputRecorder, check_model_inputs, maybe_autocast
 from .configuration_qwen3_moe import Qwen3MoeConfig
 
 
@@ -121,6 +126,7 @@ def eager_attention_forward(
     return attn_output, attn_weights
 
 
+@use_kernelized_func(apply_rotary_pos_emb)
 class Qwen3MoeAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
@@ -146,7 +152,6 @@ class Qwen3MoeAttention(nn.Module):
         self.o_proj = nn.Linear(
             config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
         )
-        self.rotary_fn = apply_rotary_pos_emb
         self.q_norm = Qwen3MoeRMSNorm(self.head_dim, eps=config.rms_norm_eps)  # unlike olmo, only on the head dim!
         self.k_norm = Qwen3MoeRMSNorm(self.head_dim, eps=config.rms_norm_eps)  # thus post q_norm does not need reshape
         self.sliding_window = getattr(config, "sliding_window", None)
@@ -212,6 +217,7 @@ class Qwen3MoeMLP(nn.Module):
         return down_proj
 
 
+@use_experts_implementation
 class Qwen3MoeExperts(nn.Module):
     """Collection of expert weights stored as 3D tensors."""
 
@@ -365,7 +371,9 @@ class Qwen3MoePreTrainedModel(PreTrainedModel):
     _supports_flash_attn = True
     _supports_sdpa = True
     _supports_flex_attn = True
-    _can_compile_fullgraph = False  # MoE models don't work with torch.compile (`torch.where(condition)` not supported)
+    _can_compile_fullgraph = (
+        is_grouped_mm_available()
+    )  # https://huggingface.co/docs/transformers/experts_interface#torchcompile
     _supports_attention_backend = True
     _can_record_outputs = {
         "router_logits": OutputRecorder(Qwen3MoeTopKRouter, layer_name="mlp.gate", index=0),
@@ -401,7 +409,7 @@ class Qwen3MoeRotaryEmbedding(nn.Module):
         inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
 
         self.register_buffer("inv_freq", inv_freq, persistent=False)
-        self.original_inv_freq = inv_freq
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
 
     @staticmethod
     def compute_default_rope_parameters(
@@ -440,7 +448,7 @@ class Qwen3MoeRotaryEmbedding(nn.Module):
         position_ids_expanded = position_ids[:, None, :].float()
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
             freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling

transformers/models/qwen3_next/modeling_qwen3_next.py

@@ -30,6 +30,7 @@ from ... import initialization as init
 from ...activations import ACT2FN
 from ...cache_utils import Cache
 from ...generation import GenerationMixin
+from ...integrations import use_experts_implementation, use_kernelized_func
 from ...masking_utils import create_causal_mask
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_layers import (
@@ -43,11 +44,8 @@ from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
 from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging
-from ...utils.generic import OutputRecorder, check_model_inputs
-from ...utils.import_utils import (
-    is_causal_conv1d_available,
-    is_flash_linear_attention_available,
-)
+from ...utils.generic import OutputRecorder, check_model_inputs, maybe_autocast
+from ...utils.import_utils import is_causal_conv1d_available, is_flash_linear_attention_available
 from .configuration_qwen3_next import Qwen3NextConfig
 
 
@@ -191,7 +189,7 @@ class Qwen3NextRotaryEmbedding(nn.Module):
         inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
 
         self.register_buffer("inv_freq", inv_freq, persistent=False)
-        self.original_inv_freq = inv_freq
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
 
     @staticmethod
     def compute_default_rope_parameters(
@@ -232,7 +230,7 @@ class Qwen3NextRotaryEmbedding(nn.Module):
         position_ids_expanded = position_ids[:, None, :].float()
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
             freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
@@ -347,6 +345,7 @@ def eager_attention_forward(
     return attn_output, attn_weights
 
 
+@use_kernelized_func(apply_rotary_pos_emb)
 class Qwen3NextAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
@@ -371,7 +370,6 @@ class Qwen3NextAttention(nn.Module):
         self.o_proj = nn.Linear(
             config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
         )
-        self.rotary_fn = apply_rotary_pos_emb
         self.q_norm = Qwen3NextRMSNorm(self.head_dim, eps=config.rms_norm_eps)  # unlike olmo, only on the head dim!
         self.k_norm = Qwen3NextRMSNorm(
             self.head_dim, eps=config.rms_norm_eps
@@ -821,6 +819,7 @@ class Qwen3NextMLP(nn.Module):
         return down_proj
 
 
+@use_experts_implementation
 class Qwen3NextExperts(nn.Module):
     """Collection of expert weights stored as 3D tensors."""
 

transformers/models/qwen3_omni_moe/configuration_qwen3_omni_moe.py

@@ -907,6 +907,7 @@ class Qwen3OmniMoeTalkerConfig(PreTrainedConfig):
         self.audio_start_token_id = audio_start_token_id
         self.vision_start_token_id = vision_start_token_id
         self.speaker_id = speaker_id
+        self.initializer_range = self.text_config.initializer_range
         super().__init__(**kwargs)
 
 
@@ -997,6 +998,7 @@ class Qwen3OmniMoeCode2WavConfig(PreTrainedConfig):
         upsampling_ratios=(2, 2),
         decoder_dim=1536,
         attention_dropout=0.0,
+        initializer_range=0.02,
         **kwargs,
     ):
         self.codebook_size = codebook_size
@@ -1016,6 +1018,7 @@ class Qwen3OmniMoeCode2WavConfig(PreTrainedConfig):
         self.upsampling_ratios = upsampling_ratios
         self.decoder_dim = decoder_dim
         self.attention_dropout = attention_dropout
+        self.initializer_range = initializer_range
         self.rope_parameters = rope_parameters
 
         super().__init__(**kwargs)
@@ -1104,6 +1107,7 @@ class Qwen3OmniMoeConfig(PreTrainedConfig):
         self.thinker_config = Qwen3OmniMoeThinkerConfig(**thinker_config)
         self.talker_config = Qwen3OmniMoeTalkerConfig(**talker_config)
         self.code2wav_config = Qwen3OmniMoeCode2WavConfig(**code2wav_config)
+        self.initializer_range = self.thinker_config.initializer_range
         self.enable_audio_output = enable_audio_output
         self.im_start_token_id = im_start_token_id
         self.im_end_token_id = im_end_token_id