[Bugfix] Fix qwen2.5-vl overflow issue (#13968)

Signed-off-by: Isotr0py <2037008807@qq.com>

vllm/model_executor/models/minicpmo.py

@@ -47,7 +47,7 @@ from .minicpmv import (MiniCPMV2_6, MiniCPMVDummyInputsBuilder,
                        MiniCPMVMultiModalDataParser,
                        MiniCPMVMultiModalProcessor, MiniCPMVProcessingInfo,
                        _minicpmv_field_config)
-from .utils import AutoWeightsLoader, maybe_prefix
+from .utils import AutoWeightsLoader, cast_overflow_tensors, maybe_prefix
 
 CPU_DEVICE = torch.device("cpu")
 
@@ -469,13 +469,8 @@ class MiniCPMWhisperEncoderLayer(nn.Module):
                                               training=self.training)
         hidden_states = residual + hidden_states
 
-        if hidden_states.dtype == torch.float16 and (
-                torch.isinf(hidden_states).any()
-                or torch.isnan(hidden_states).any()):
-            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
-            hidden_states = torch.clamp(hidden_states,
-                                        min=-clamp_value,
-                                        max=clamp_value)
+        if hidden_states.dtype == torch.float16:
+            hidden_states = cast_overflow_tensors(hidden_states)
 
         outputs = (hidden_states, )
 

vllm/model_executor/models/qwen2_5_vl.py

@@ -63,7 +63,7 @@ from .interfaces import SupportsLoRA, SupportsMultiModal, SupportsPP
 from .qwen2_vl import Qwen2VLDummyInputsBuilder as Qwen2_5_VLDummyInputsBuilder
 from .qwen2_vl import (Qwen2VLMultiModalProcessor, Qwen2VLProcessingInfo,
                        apply_rotary_pos_emb_vision)
-from .utils import (AutoWeightsLoader, WeightsMapper,
+from .utils import (AutoWeightsLoader, WeightsMapper, cast_overflow_tensors,
                     init_vllm_registered_model, maybe_prefix,
                     merge_multimodal_embeddings)
 from .vision import get_vit_attn_backend
@@ -641,6 +641,11 @@ class Qwen2_5_VisionTransformer(nn.Module):
                                 cu_seqlens=cu_seqlens_now,
                                 rotary_pos_emb=rotary_pos_emb)
 
+        # For Qwen2.5-VL-3B, float16 will overflow at last block
+        # for long visual tokens sequences.
+        if hidden_states.dtype == torch.float16:
+            hidden_states = cast_overflow_tensors(hidden_states)
+
         # adapter
         hidden_states = self.merger(hidden_states)
         reverse_indices = torch.argsort(window_index)

vllm/model_executor/models/utils.py

@@ -641,3 +641,13 @@ def extract_layer_index(layer_name: str) -> int:
     assert len(int_vals) == 1, (f"layer name {layer_name} should"
                                 " only contain one integer")
     return int_vals[0]
+
+
+def cast_overflow_tensors(
+    tensors: torch.Tensor,
+    offset: float = 1000,
+) -> torch.Tensor:
+    if tensors.isinf().any() or tensors.isnan().any():
+        clamp_value = torch.finfo(tensors.dtype).max - offset
+        tensors = torch.clamp(tensors, min=-clamp_value, max=clamp_value)
+    return tensors

vllm/model_executor/models/whisper.py

@@ -35,7 +35,8 @@ from vllm.multimodal.processing import (BaseProcessingInfo,
 from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
 
 from .interfaces import SupportsMultiModal, SupportsTranscription
-from .utils import AutoWeightsLoader, WeightsMapper, make_layers
+from .utils import (AutoWeightsLoader, WeightsMapper, cast_overflow_tensors,
+                    make_layers)
 
 logger = init_logger(__name__)
 
@@ -285,11 +286,7 @@ class WhisperEncoderLayer(nn.Module):
         hidden_states = self.mlp(hidden_states)
         hidden_states = residual + hidden_states
 
-        if hidden_states.isinf().any() or hidden_states.isnan().any():
-            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
-            hidden_states = torch.clamp(hidden_states,
-                                        min=-clamp_value,
-                                        max=clamp_value)
+        hidden_states = cast_overflow_tensors(hidden_states)
 
         return hidden_states