[Kernel] Support CUDA Graphs in 3D Triton Attention Kernel (#28306)

Signed-off-by: Jan van Lunteren <jvl@zurich.ibm.com>
Signed-off-by: jvlunteren <161835099+jvlunteren@users.noreply.github.com>
Co-authored-by: Thomas Parnell <tom.parnell@gmail.com>
Co-authored-by: Thomas Parnell <tpa@zurich.ibm.com>

tests/kernels/attention/test_triton_unified_attention.py

@@ -7,6 +7,7 @@ import torch
 
 from vllm.attention.ops.triton_unified_attention import unified_attention
 from vllm.platforms import current_platform
+from vllm.utils.math_utils import next_power_of_2
 
 NUM_HEADS = [(4, 4), (8, 2)]
 HEAD_SIZES = [128, 256]
@@ -22,6 +23,10 @@ QDTYPES = (
 # one value small enough to test the schema op check
 NUM_BLOCKS = [32768, 2048]
 
+# 0: use 2D kernel for decode
+# 8: use 3D kernel for decode
+SEQ_THRESHOLD_3D_VALUES = [0, 8]
+
 
 def ref_paged_attn(
     query: torch.Tensor,
@@ -92,6 +97,7 @@ def ref_paged_attn(
 @pytest.mark.parametrize("soft_cap", [None, 50.0])
 @pytest.mark.parametrize("num_blocks", NUM_BLOCKS)
 @pytest.mark.parametrize("q_dtype", QDTYPES)
+@pytest.mark.parametrize("seq_threshold_3D", SEQ_THRESHOLD_3D_VALUES)
 @torch.inference_mode()
 def test_triton_unified_attn(
     seq_lens: list[tuple[int, int]],
@@ -103,6 +109,7 @@ def test_triton_unified_attn(
     soft_cap: float | None,
     num_blocks: int,
     q_dtype: torch.dtype | None,
+    seq_threshold_3D: int,
 ) -> None:
     torch.set_default_device("cuda")
 
@@ -152,6 +159,21 @@ def test_triton_unified_attn(
         k_descale = torch.rand(scale_shape, dtype=torch.float32)
         v_descale = torch.rand(scale_shape, dtype=torch.float32)
 
+    num_par_softmax_segments = 16
+    head_size_padded = next_power_of_2(head_size)
+    softmax_segm_output = torch.empty(
+        (seq_threshold_3D, num_query_heads, num_par_softmax_segments, head_size_padded),
+        dtype=torch.float32,
+    )
+    softmax_segm_max = torch.empty(
+        (seq_threshold_3D, num_query_heads, num_par_softmax_segments),
+        dtype=torch.float32,
+    )
+    softmax_segm_expsum = torch.empty(
+        (seq_threshold_3D, num_query_heads, num_par_softmax_segments),
+        dtype=torch.float32,
+    )
+
     unified_attention(
         q=maybe_quantized_query,
         k=maybe_quantized_key_cache,
@@ -169,6 +191,11 @@ def test_triton_unified_attn(
         q_descale=q_descale,
         k_descale=k_descale,
         v_descale=v_descale,
+        seq_threshold_3D=seq_threshold_3D,
+        num_par_softmax_segments=num_par_softmax_segments,
+        softmax_segm_output=softmax_segm_output,
+        softmax_segm_max=softmax_segm_max,
+        softmax_segm_expsum=softmax_segm_expsum,
     )
 
     ref_output = ref_paged_attn(

vllm/attention/ops/triton_unified_attention.py

@@ -355,7 +355,7 @@ def kernel_unified_attention_2d(
 @triton.jit
 def kernel_unified_attention_3d(
     segm_output_ptr,
-    # [num_tokens, num_query_heads, num_segments, head_size]
+    # [num_tokens, num_query_heads, num_segments, head_size_padded]
     segm_max_ptr,  # [num_tokens, num_query_heads, num_segments]
     segm_expsum_ptr,  # [num_tokens, num_query_heads, num_segments]
     query_ptr,  # [num_tokens, num_query_heads, head_size]
@@ -749,6 +749,11 @@ def unified_attention(
     q_descale,
     k_descale,
     v_descale,
+    seq_threshold_3D=None,
+    num_par_softmax_segments=None,
+    softmax_segm_output=None,
+    softmax_segm_max=None,
+    softmax_segm_expsum=None,
     alibi_slopes=None,
     output_scale=None,
     qq_bias=None,
@@ -793,8 +798,19 @@ def unified_attention(
     TILE_SIZE_PREFILL = 32
     TILE_SIZE_DECODE = 16 if q.element_size() >= 2 else 32
 
-    # if batch contains a prefill
-    if max_seqlen_q > 1 or total_num_q_blocks * num_kv_heads > 128:
+    # Launch the 2D kernel if
+    # 1. No intermediate tiled softmax buffers for the 3D kernel have been allocated, or
+    # 2. The batch includes at least one prefill request, or
+    # 3. The number of sequences exceeds the configured threshold
+    if (
+        seq_threshold_3D is None
+        or num_par_softmax_segments is None
+        or softmax_segm_output is None
+        or softmax_segm_max is None
+        or softmax_segm_expsum is None
+        or max_seqlen_q > 1
+        or num_seqs > seq_threshold_3D
+    ):
         kernel_unified_attention_2d[
             (
                 total_num_q_blocks,
@@ -847,37 +863,12 @@ def unified_attention(
             USE_FP8=output_scale is not None,
         )
     else:
-        # for initial version, NUM_SEGMENTS = 16 is chosen as a default
-        # value that showed good performance in tests
-        NUM_SEGMENTS = 16
-
-        segm_output = torch.empty(
-            q.shape[0],
-            num_query_heads,
-            NUM_SEGMENTS,
-            triton.next_power_of_2(head_size),
-            dtype=torch.float32,
-            device=q.device,
-        )
-        segm_max = torch.empty(
-            q.shape[0],
-            num_query_heads,
-            NUM_SEGMENTS,
-            dtype=torch.float32,
-            device=q.device,
-        )
-        segm_expsum = torch.empty(
-            q.shape[0],
-            num_query_heads,
-            NUM_SEGMENTS,
-            dtype=torch.float32,
-            device=q.device,
-        )
-
-        kernel_unified_attention_3d[(total_num_q_blocks, num_kv_heads, NUM_SEGMENTS)](
-            segm_output_ptr=segm_output,
-            segm_max_ptr=segm_max,
-            segm_expsum_ptr=segm_expsum,
+        kernel_unified_attention_3d[
+            (total_num_q_blocks, num_kv_heads, num_par_softmax_segments)
+        ](
+            segm_output_ptr=softmax_segm_output,
+            segm_max_ptr=softmax_segm_max,
+            segm_expsum_ptr=softmax_segm_expsum,
             query_ptr=q,
             key_cache_ptr=k,
             value_cache_ptr=v,
@@ -917,13 +908,13 @@ def unified_attention(
             BLOCK_Q=BLOCK_Q,
             num_seqs=num_seqs,
             BLOCK_M=BLOCK_M,
-            NUM_SEGMENTS_PER_SEQ=NUM_SEGMENTS,
+            NUM_SEGMENTS_PER_SEQ=num_par_softmax_segments,
         )
         reduce_segments[(q.shape[0], num_query_heads)](
             output_ptr=out,
-            segm_output_ptr=segm_output,
-            segm_max_ptr=segm_max,
-            segm_expsum_ptr=segm_expsum,
+            segm_output_ptr=softmax_segm_output,
+            segm_max_ptr=softmax_segm_max,
+            segm_expsum_ptr=softmax_segm_expsum,
             seq_lens_ptr=seqused_k,
             num_seqs=num_seqs,
             num_query_heads=num_query_heads,
@@ -936,6 +927,6 @@ def unified_attention(
             HEAD_SIZE_PADDED=triton.next_power_of_2(head_size),
             query_start_len_ptr=cu_seqlens_q,
             BLOCK_Q=BLOCK_Q,
-            NUM_SEGMENTS_PER_SEQ=NUM_SEGMENTS,
+            NUM_SEGMENTS_PER_SEQ=num_par_softmax_segments,
             USE_FP8=output_scale is not None,
         )

vllm/v1/attention/backends/triton_attn.py

@@ -17,7 +17,7 @@ from vllm.attention.ops.triton_reshape_and_cache_flash import (
     triton_reshape_and_cache_flash,
 )
 from vllm.attention.ops.triton_unified_attention import unified_attention
-from vllm.config import VllmConfig
+from vllm.config import CUDAGraphMode, VllmConfig
 from vllm.config.cache import CacheDType
 from vllm.logger import init_logger
 from vllm.model_executor.layers.quantization.utils.quant_utils import (
@@ -26,6 +26,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
 )
 from vllm.platforms import current_platform
 from vllm.platforms.interface import DeviceCapability
+from vllm.utils.math_utils import next_power_of_2
 from vllm.v1.attention.backends.utils import (
     AttentionCGSupport,
     AttentionMetadataBuilder,
@@ -36,6 +37,11 @@ from vllm.v1.kv_cache_interface import AttentionSpec
 logger = init_logger(__name__)
 
 
+# constants
+MIN_LAUNCH_GRID_SIZE_2D = 128  # Minimum launch grid size of 2D kernel
+NUM_PAR_SOFTMAX_SEGMENTS = 16  # Number of parallel tiled softmax segments
+
+
 @dataclass
 class TritonAttentionMetadata:
     # NOTE(sang): Definition of context_len, query_len, and seq_len.
@@ -54,6 +60,12 @@ class TritonAttentionMetadata:
     block_table: torch.Tensor
     slot_mapping: torch.Tensor
 
+    seq_threshold_3D: int
+    num_par_softmax_segments: int
+    softmax_segm_output: torch.Tensor
+    softmax_segm_max: torch.Tensor
+    softmax_segm_expsum: torch.Tensor
+
     # For cascade attention.
     use_cascade: bool
     common_prefix_len: int
@@ -87,6 +99,60 @@ class TritonAttentionMetadataBuilder(AttentionMetadataBuilder[TritonAttentionMet
         self.num_heads_kv = model_config.get_num_kv_heads(vllm_config.parallel_config)
         self.headdim = model_config.get_head_size()
 
+        # Check if CUDA Graphs are enabled for decode
+        self.decode_cudagraph_enabled = (
+            self.vllm_config.compilation_config.cudagraph_mode
+            in (
+                CUDAGraphMode.FULL_AND_PIECEWISE,
+                CUDAGraphMode.FULL_DECODE_ONLY,
+                CUDAGraphMode.FULL,
+            )
+        )
+
+        # The launch grid for the 2D kernel is defined as (num_q_blocks, num_heads_kv).
+        # A lower bound for num_q_blocks is the number of sequences.
+        # To ensure the minimum launch grid size is achieved, the number of sequences
+        # must be at least equal to the threshold below.
+        # If this threshold is not reached (i.e., the batch size is not large enough),
+        # the 3D kernel will be selected instead.
+        self.seq_threshold_3D = MIN_LAUNCH_GRID_SIZE_2D // self.num_heads_kv
+
+        # Modify the threshold if needed.
+        if self.decode_cudagraph_enabled:
+            capture_sizes = self.vllm_config.compilation_config.cudagraph_capture_sizes
+            assert capture_sizes, "CUDA Graphs enabled but no capture sizes specified."
+
+            # Select the CUDA Graph capture size closest to self.seq_threshold_3D
+            # as threshold. This ensures that each captured graph covers the
+            # correct execution path.
+            self.seq_threshold_3D = min(
+                capture_sizes,
+                key=lambda x: abs(x - self.seq_threshold_3D),
+            )
+
+        self.num_par_softmax_segments = NUM_PAR_SOFTMAX_SEGMENTS
+        headdim_padded = next_power_of_2(self.headdim)
+        self.softmax_segm_output = torch.empty(
+            (
+                self.seq_threshold_3D,
+                self.num_heads_q,
+                self.num_par_softmax_segments,
+                headdim_padded,
+            ),
+            dtype=torch.float32,
+            device=device,
+        )
+        self.softmax_segm_max = torch.empty(
+            (self.seq_threshold_3D, self.num_heads_q, self.num_par_softmax_segments),
+            dtype=torch.float32,
+            device=device,
+        )
+        self.softmax_segm_expsum = torch.empty(
+            (self.seq_threshold_3D, self.num_heads_q, self.num_par_softmax_segments),
+            dtype=torch.float32,
+            device=device,
+        )
+
     def build_for_cudagraph_capture(
         self, common_attn_metadata: CommonAttentionMetadata
     ) -> TritonAttentionMetadata:
@@ -143,6 +209,11 @@ class TritonAttentionMetadataBuilder(AttentionMetadataBuilder[TritonAttentionMet
             prefix_kv_lens=prefix_kv_lens,
             suffix_kv_lens=suffix_kv_lens,
             prefix_scheduler_metadata=prefix_scheduler_metadata,
+            seq_threshold_3D=self.seq_threshold_3D,
+            num_par_softmax_segments=self.num_par_softmax_segments,
+            softmax_segm_output=self.softmax_segm_output,
+            softmax_segm_max=self.softmax_segm_max,
+            softmax_segm_expsum=self.softmax_segm_expsum,
         )
         return attn_metadata
 
@@ -349,6 +420,12 @@ class TritonAttentionImpl(AttentionImpl):
         max_seqlen_k = attn_metadata.max_seq_len
         block_table = attn_metadata.block_table
 
+        seq_threshold_3D = attn_metadata.seq_threshold_3D
+        num_par_softmax_segments = attn_metadata.num_par_softmax_segments
+        softmax_segm_output = attn_metadata.softmax_segm_output
+        softmax_segm_max = attn_metadata.softmax_segm_max
+        softmax_segm_expsum = attn_metadata.softmax_segm_expsum
+
         descale_shape = (cu_seqlens_q.shape[0] - 1, key_cache.shape[2])
 
         unified_attention(
@@ -369,6 +446,11 @@ class TritonAttentionImpl(AttentionImpl):
             q_descale=None,  # Not supported
             k_descale=layer._k_scale.expand(descale_shape),
             v_descale=layer._v_scale.expand(descale_shape),
+            seq_threshold_3D=seq_threshold_3D,
+            num_par_softmax_segments=num_par_softmax_segments,
+            softmax_segm_output=softmax_segm_output,
+            softmax_segm_max=softmax_segm_max,
+            softmax_segm_expsum=softmax_segm_expsum,
             sinks=self.sinks,
             output_scale=output_scale,
         )