[Fix]Fix capture fail bug for DeepSeek (#6275)

python/sglang/srt/mem_cache/memory_pool.py

@@ -266,7 +266,6 @@ class MHATokenToKVPool(KVCache):
         self._create_buffers()
 
         self.layer_transfer_counter = None
-        self.capture_mode = False
         self.device_module = torch.get_device_module(self.device)
         self.alt_stream = self.device_module.Stream() if is_cuda else None
 
@@ -385,6 +384,8 @@ class MHATokenToKVPool(KVCache):
         k_scale: Optional[float] = None,
         v_scale: Optional[float] = None,
     ):
+        from sglang.srt.model_executor.cuda_graph_runner import get_is_capture_mode
+
         layer_id = layer.layer_id
         if cache_k.dtype != self.dtype:
             if k_scale is not None:
@@ -398,7 +399,7 @@ class MHATokenToKVPool(KVCache):
             cache_k = cache_k.view(self.store_dtype)
             cache_v = cache_v.view(self.store_dtype)
 
-        if self.capture_mode and self.alt_stream is not None:
+        if get_is_capture_mode() and self.alt_stream is not None:
             # Overlap the copy of K and V cache for small batch size
             current_stream = self.device_module.current_stream()
             self.alt_stream.wait_stream(current_stream)

python/sglang/srt/model_executor/cuda_graph_runner.py

@@ -47,6 +47,13 @@ from sglang.srt.utils import (
 if TYPE_CHECKING:
     from sglang.srt.model_executor.model_runner import ModelRunner
 
+# Detect whether the current forward pass is in capture mode
+is_capture_mode = False
+
+
+def get_is_capture_mode():
+    return is_capture_mode
+
 
 def _to_torch(model: torch.nn.Module, reverse: bool, num_tokens: int):
     for sub in model._modules.values():
@@ -311,17 +318,12 @@ class CudaGraphRunner:
 
     @contextmanager
     def model_capture_mode(self):
-        if hasattr(self.model_runner.model, "capture_mode"):
-            self.model_runner.model.capture_mode = True
-        if hasattr(self.model_runner.token_to_kv_pool, "capture_mode"):
-            self.model_runner.token_to_kv_pool.capture_mode = True
+        global is_capture_mode
+        is_capture_mode = True
 
         yield
 
-        if hasattr(self.model_runner.model, "capture_mode"):
-            self.model_runner.model.capture_mode = False
-        if hasattr(self.model_runner.token_to_kv_pool, "capture_mode"):
-            self.model_runner.token_to_kv_pool.capture_mode = False
+        is_capture_mode = False
 
     def can_run(self, forward_batch: ForwardBatch):
         if self.enable_dp_attention or self.enable_sp_layernorm:
@@ -612,6 +614,7 @@ class CudaGraphRunner:
 
         # Replay
         self.graphs[self.bs].replay()
+
         output = self.output_buffers[self.bs]
         if isinstance(output, LogitsProcessorOutput):
             return LogitsProcessorOutput(

python/sglang/srt/models/deepseek_v2.py

@@ -754,6 +754,8 @@ class DeepseekV2AttentionMLA(nn.Module):
         forward_batch: ForwardBatch,
         zero_allocator: BumpAllocator,
     ) -> torch.Tensor:
+        from sglang.srt.model_executor.cuda_graph_runner import get_is_capture_mode
+
         if self.q_lora_rank is not None:
             q, latent_cache = self.fused_qkv_a_proj_with_mqa(hidden_states)[0].split(
                 [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim], dim=-1
@@ -761,7 +763,7 @@ class DeepseekV2AttentionMLA(nn.Module):
             k_nope = latent_cache[..., : self.kv_lora_rank]
 
             # overlap qk norm
-            if self.alt_stream is not None and torch.cuda.is_current_stream_capturing():
+            if self.alt_stream is not None and get_is_capture_mode():
                 current_stream = torch.cuda.current_stream()
                 self.alt_stream.wait_stream(current_stream)
                 q = self.q_a_layernorm(q)

python/sglang/srt/models/mllama.py

@@ -836,7 +836,6 @@ class MllamaForConditionalGeneration(nn.Module):
             prefix="multi_modal_projector",
         )
         self.logits_processor = LogitsProcessor(config.text_config)
-        self.capture_mode = False
 
     def pad_input_ids(self, input_ids: List[int], mm_inputs: MultimodalInputs):
         pixel_values = torch.cat(
@@ -969,6 +968,8 @@ class MllamaForConditionalGeneration(nn.Module):
         positions: torch.Tensor,
         forward_batch: ForwardBatch,
     ) -> Union[Tuple, CausalLMOutputWithPast]:
+        from sglang.srt.model_executor.cuda_graph_runner import get_is_capture_mode
+
         batched_images, batched_ar_ids, batched_ar_mask, encoder_lens_need = (
             self._batch_image_inputs(forward_batch)
         )
@@ -977,7 +978,7 @@ class MllamaForConditionalGeneration(nn.Module):
         cross_attention_mask = None
         cross_attention_states = None
 
-        if self.capture_mode:
+        if get_is_capture_mode():
             # NOTE: when doing cuda graph capture, we do not want to skip cross attention
             # Make is a constant value to avoid cuda graph capture issue
             skip_cross_attention = False