[Nvidia] Integrate SM100 cudnn prefill API to MLA prefill (#20411)

Signed-off-by: Elfie Guo <elfieg@nvidia.com>
Co-authored-by: Elfie Guo <eflieg@nvidia.com>

vllm/envs.py

@@ -139,6 +139,7 @@ if TYPE_CHECKING:
     VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16: bool = True
     VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB: Optional[int] = None
     VLLM_NIXL_ABORT_REQUEST_TIMEOUT: int = 120
+    VLLM_USE_CUDNN_PREFILL: bool = False
     VLLM_LOOPBACK_IP: str = ""
 
 
@@ -962,6 +963,10 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "VLLM_NIXL_ABORT_REQUEST_TIMEOUT":
     lambda: int(os.getenv("VLLM_NIXL_ABORT_REQUEST_TIMEOUT", "120")),
 
+    # Controls whether or not to use cudnn prefill
+    "VLLM_USE_CUDNN_PREFILL":
+    lambda: bool(int(os.getenv("VLLM_USE_CUDNN_PREFILL", "0"))),
+
     # If set to 1, use the TRTLLM Decode Attention backend in flashinfer.
     "VLLM_USE_TRTLLM_DECODE_ATTENTION":
     lambda: os.getenv("VLLM_USE_TRTLLM_DECODE_ATTENTION", None),

vllm/v1/attention/backends/mla/common.py

@@ -194,6 +194,7 @@ from typing import TYPE_CHECKING, Any, Generic, Optional, TypeVar, Union
 
 import torch
 
+import vllm.envs as envs
 from vllm import _custom_ops as ops
 from vllm.attention.backends.abstract import (AttentionBackend, AttentionLayer,
                                               AttentionMetadata,
@@ -225,6 +226,8 @@ except ImportError:
 
 try:
     from flashinfer import BatchPrefillWithRaggedKVCacheWrapper
+    from flashinfer.prefill import (  # noqa: F401
+        cudnn_batch_prefill_with_kv_cache)
     flashinfer_available = True
 except ImportError:
     flashinfer_available = False
@@ -236,6 +239,8 @@ if TYPE_CHECKING:
 
 logger = init_logger(__name__)
 
+CUDNN_WORKSPACE_SIZE = 12800
+
 
 class MLACommonBackend(AttentionBackend):
 
@@ -294,6 +299,7 @@ class MLACommonPrefillMetadata:
         starts: torch.Tensor
         seq_tot: list[int]
         max_seq_lens: list[int]
+        seq_lens: torch.Tensor
         workspace: torch.Tensor
 
     block_table: torch.Tensor
@@ -309,6 +315,17 @@ class FlashInferPrefillMetadata(MLACommonPrefillMetadata):
         default_factory=list)
 
 
+@dataclass
+class CudnnPrefillMetadata(MLACommonPrefillMetadata):
+
+    class ChunkedContextMetadata(
+            MLACommonPrefillMetadata.ChunkedContextMetadata):
+        seq_lens: torch.Tensor
+
+    query_seq_lens: Optional[torch.Tensor] = None
+    cudnn_workspace: Optional[torch.Tensor] = None
+
+
 @dataclass
 class MLACommonDecodeMetadata:
     block_table: torch.Tensor
@@ -351,7 +368,8 @@ class MLACommonMetadata(Generic[D]):
 
     decode: Optional[D] = None
     prefill: Optional[Union[MLACommonPrefillMetadata,
-                            FlashInferPrefillMetadata]] = None
+                            FlashInferPrefillMetadata,
+                            CudnnPrefillMetadata]] = None
 
     def __post_init__(self):
         if self.head_dim is not None:
@@ -362,13 +380,19 @@ M = TypeVar("M", bound=MLACommonMetadata)
 
 
 def use_flashinfer_prefill() -> bool:
-    if flashinfer_available:
+    if flashinfer_available and not envs.VLLM_USE_CUDNN_PREFILL:
         # For blackwell default to flashinfer prefill if its available since
         #  its faster than FA2.
         return current_platform.has_device_capability(100)
     return False
 
 
+def use_cudnn_prefill() -> bool:
+    if flashinfer_available and envs.VLLM_USE_CUDNN_PREFILL:
+        return current_platform.has_device_capability(100)
+    return False
+
+
 # Currently 394MB, this can be tuned based on GEMM sizes used.
 # Choosen to be the same as sglang:
 #  https://github.com/sgl-project/sglang/blob/766392c6bda2558b61ce6d1c1bfd8081a549e1f1/python/sglang/global_config.py#L37
@@ -427,11 +451,15 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
                 dtype=model_config.dtype,
                 device=runner.device,
             )
+
         self.block_table = block_table
 
+        self._use_cudnn_prefill = use_cudnn_prefill()
         self._use_fi_prefill = use_flashinfer_prefill()
-        self.prefill_metadata_cls = FlashInferPrefillMetadata \
-            if self._use_fi_prefill else MLACommonPrefillMetadata
+        self.prefill_metadata_cls = (
+            FlashInferPrefillMetadata
+            if self._use_fi_prefill else CudnnPrefillMetadata
+            if self._use_cudnn_prefill else MLACommonPrefillMetadata)
 
         if self._use_fi_prefill:
             self._workspace_buffer = torch.empty(
@@ -447,6 +475,13 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
             self._global_hyperparameters = infer_global_hyperparameters(
                 get_per_layer_parameters(runner.vllm_config, MLACommonImpl))
 
+        if self._use_cudnn_prefill:
+            self.cudnn_workspace = torch.empty(
+                CUDNN_WORKSPACE_SIZE * scheduler_config.max_num_seqs,
+                dtype=torch.int8,
+                device=runner.device,
+            )
+
     def _build_fi_prefill_wrappers(self, prefill: FlashInferPrefillMetadata):
         qo_indptr = prefill.query_start_loc
 
@@ -692,15 +727,24 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
                              out=cu_seq_lens_cpu[:, 1:],
                              dtype=torch.int32)
 
+                chunked_context_metadata_cls = \
+                    CudnnPrefillMetadata.ChunkedContextMetadata \
+                    if self._use_cudnn_prefill else \
+                        MLACommonPrefillMetadata.ChunkedContextMetadata
+
                 chunked_context_metadata = \
-                    MLACommonPrefillMetadata.ChunkedContextMetadata(
+                    chunked_context_metadata_cls(
                     cu_seq_lens=cu_seq_lens_cpu.to(device, non_blocking=True),
                     starts=chunk_starts.to(device, non_blocking=True),
                     seq_tot=chunk_seq_lens.sum(dim=1).tolist(),
                     max_seq_lens=chunk_seq_lens.max(dim=1).values.tolist(),
+                    seq_lens=chunk_seq_lens,
                     workspace=self.chunked_prefill_workspace,
                 )
 
+                if self._use_cudnn_prefill:
+                    chunked_context_metadata.seq_lens = chunk_seq_lens
+
                 assert max(chunked_context_metadata.max_seq_lens) <= \
                     self.chunked_prefill_workspace_size
 
@@ -711,6 +755,12 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
                 chunked_context=chunked_context_metadata,
             )
 
+            if self._use_cudnn_prefill:
+                assert isinstance(prefill_metadata, CudnnPrefillMetadata)
+                prefill_metadata.query_seq_lens = prefill_query_start_loc[1:] \
+                    - prefill_query_start_loc[:-1]
+                prefill_metadata.cudnn_workspace = self.cudnn_workspace
+
         decode_metadata = None
         if self._num_decodes > 0:
             decode_metadata = self._build_decode(
@@ -794,6 +844,12 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
             self._run_prefill_context_chunk = self._run_prefill_context_chunk_fi
             self._run_prefill_new_tokens = self._run_prefill_new_tokens_fi
             self._pad_v = False
+        elif use_cudnn_prefill():
+            logger.debug_once("Using CUDNN prefill for MLA")
+            self._run_prefill_context_chunk = \
+                self._run_prefill_context_chunk_cudnn
+            self._run_prefill_new_tokens = self._run_prefill_new_tokens_cudnn
+            self._pad_v = False
         else:  # Use FlashAttention
             logger.debug_once("Using FlashAttention prefill for MLA")
             self._run_prefill_context_chunk = self._run_prefill_context_chunk_fa
@@ -882,6 +938,29 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
             return_lse=return_softmax_lse,
         )
 
+    def _run_prefill_new_tokens_cudnn(self, prefill: MLACommonPrefillMetadata,
+                                      q, k, v, return_softmax_lse):
+        assert isinstance(prefill, CudnnPrefillMetadata)
+        assert prefill.query_seq_lens is not None
+        output, lse = cudnn_batch_prefill_with_kv_cache(
+            q=q,
+            k_cache=k,
+            v_cache=v,
+            scale=self.scale,
+            workspace_buffer=prefill.cudnn_workspace,
+            max_token_per_sequence=prefill.max_query_len,
+            max_sequence_kv=prefill.max_query_len,
+            actual_seq_lens_q=prefill.query_seq_lens.view(-1, 1, 1, 1),
+            actual_seq_lens_kv=prefill.query_seq_lens.view(-1, 1, 1, 1),
+            causal=True,
+            return_lse=True,  # do not support False for now
+            is_cuda_graph_compatible=
+            True,  #Indicates actual_seq_lens are on GPU or CPU.
+        )
+        if return_softmax_lse:
+            return output, lse
+        return output
+
     def _run_prefill_context_chunk_fa(self, prefill: MLACommonPrefillMetadata,
                                       chunk_idx: int, q, k, v):
         assert prefill.chunked_context is not None
@@ -908,6 +987,30 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
             return_lse=True,
         )
 
+    def _run_prefill_context_chunk_cudnn(self,
+                                         prefill: MLACommonPrefillMetadata,
+                                         chunk_idx: int, q, k, v):
+        assert isinstance(prefill, CudnnPrefillMetadata)
+        assert prefill.chunked_context is not None
+        assert prefill.chunked_context.seq_lens[chunk_idx] is not None
+        assert prefill.query_seq_lens is not None
+        return cudnn_batch_prefill_with_kv_cache(
+            q=q,
+            k_cache=k,
+            v_cache=v,
+            scale=self.scale,
+            workspace_buffer=prefill.cudnn_workspace,
+            max_token_per_sequence=prefill.max_query_len,
+            max_sequence_kv=prefill.chunked_context.max_seq_lens[chunk_idx],
+            actual_seq_lens_q=prefill.query_seq_lens.view(-1, 1, 1, 1),
+            actual_seq_lens_kv=prefill.chunked_context.seq_lens[chunk_idx].
+            view(-1, 1, 1, 1),
+            causal=False,
+            return_lse=True,
+            is_cuda_graph_compatible=
+            True,  #Indicates actual_seq_lens are on GPU or CPU.
+        )
+
     def _v_up_proj(self, x):
         # Convert from (B, N, L) to (N, B, L)
         x = x.view(-1, self.num_heads, self.kv_lora_rank).transpose(0, 1)