[Flashinfer][gpt-oss] Support FP8-qkv Flashinfer TRTLLM Sinks Attention (#25674)

Signed-off-by: elvischenv <219235043+elvischenv@users.noreply.github.com>

tests/kernels/attention/test_flashinfer_trtllm_attention.py

@@ -49,6 +49,7 @@ KV_LAYOUT = ["HND"]  # currently only HND is supported
 BLOCK_SIZE = [16]
 WINDOW_LEFT = [-1, 127]
 SOFT_CAP = [None, 50.0]
+HAS_SINKS = [True, False]
 
 NUM_BLOCKS = 32768  # Large enough to test overflow in index calculation.
 
@@ -63,6 +64,7 @@ NUM_BLOCKS = 32768  # Large enough to test overflow in index calculation.
 @pytest.mark.parametrize("block_size", BLOCK_SIZE)
 @pytest.mark.parametrize("window_left", WINDOW_LEFT)
 @pytest.mark.parametrize("soft_cap", SOFT_CAP)
+@pytest.mark.parametrize("has_sinks", HAS_SINKS)
 @torch.inference_mode
 def test_flashinfer_trtllm_decode_with_baseline(
     dtype: torch.dtype,
@@ -77,9 +79,10 @@ def test_flashinfer_trtllm_decode_with_baseline(
     block_size: int,
     window_left: int,
     soft_cap: Optional[float],
+    has_sinks: bool,
 ) -> None:
     torch.set_default_device("cuda")
-    current_platform.seed_everything(0)
+    current_platform.seed_everything(42)
 
     q_quant_dtype, kv_quant_dtype, o_quant_dtype = quant_dtypes
     q_quant_dtype = q_quant_dtype or dtype
@@ -101,7 +104,16 @@ def test_flashinfer_trtllm_decode_with_baseline(
     else:
         raise ValueError(f"Invalid kv_layout: {kv_layout}")
 
-    query = torch.randn(batch_size, num_qo_heads, head_size, dtype=dtype)
+    # max_q_len = 1
+    q_lens = torch.ones((batch_size,), dtype=torch.int32)
+    q_indptr = torch.cat(
+        [
+            torch.tensor([0], dtype=torch.int32),
+            torch.cumsum(q_lens, dim=0, dtype=torch.int32),
+        ]
+    )
+
+    query = torch.randn(torch.sum(q_lens).item(), num_qo_heads, head_size, dtype=dtype)
     if q_quant_dtype == FP8_DTYPE:
         query, q_scale = to_float8(query)
         ref_query = query.to(dtype) * q_scale
@@ -112,7 +124,7 @@ def test_flashinfer_trtllm_decode_with_baseline(
     kv_lens = torch.randint(1, max_kv_len, (batch_size,), dtype=torch.int32)
     kv_lens[-1] = max_kv_len
 
-    seq_lens = kv_lens
+    seq_lens = kv_lens + q_lens
     max_seq_len = torch.max(seq_lens).item()
 
     kv_cache = torch.randn(kv_cache_shape, dtype=dtype)
@@ -148,27 +160,36 @@ def test_flashinfer_trtllm_decode_with_baseline(
     workspace_buffer = torch.zeros(128 * 1024 * 1024, dtype=torch.int8)
 
     # Baseline Decode
-    wrapper = flashinfer.BatchDecodeWithPagedKVCacheWrapper(
-        workspace_buffer, kv_layout, use_tensor_cores=True
+    if has_sinks:
+        sinks = torch.rand(num_qo_heads, dtype=torch.float32) * 5
+        wrapper = flashinfer.BatchAttentionWithAttentionSinkWrapper(
+            float_workspace_buffer=workspace_buffer, kv_layout=kv_layout, backend="fa2"
         )
+    else:
+        sinks = None
+        wrapper = flashinfer.BatchPrefillWithPagedKVCacheWrapper(
+            float_workspace_buffer=workspace_buffer, kv_layout=kv_layout, backend="fa2"
+        )
+
     wrapper.plan(
-        kv_indptr,
-        kv_indices,
-        kv_last_page_lens,
-        num_qo_heads,
-        num_kv_heads,
-        head_size,
-        block_size,
-        "NONE",
+        qo_indptr=q_indptr,
+        paged_kv_indptr=kv_indptr,
+        paged_kv_indices=kv_indices,
+        paged_kv_last_page_len=kv_last_page_lens,
+        num_qo_heads=num_qo_heads,
+        num_kv_heads=num_kv_heads,
+        head_dim_qk=head_size,
+        page_size=block_size,
+        causal=True,
         sm_scale=sm_scale,
-        q_data_type=dtype,
-        kv_data_type=dtype,
         window_left=window_left,
         logits_soft_cap=soft_cap,
+        q_data_type=dtype,
+        kv_data_type=dtype,
     )
-
     output = torch.empty(ref_query.shape, dtype=dtype)
-    wrapper.run(ref_query, ref_kv_cache, out=output)
+    wrapper.run(ref_query, ref_kv_cache, sinks, sm_scale, out=output)
+
     o_scale = 1.0
     o_sf_scale_float = None
     if o_quant_dtype == FP8_DTYPE:
@@ -202,6 +223,7 @@ def test_flashinfer_trtllm_decode_with_baseline(
         bmm1_scale=q_scale * k_scale * sm_scale,
         bmm2_scale=v_scale / o_scale,
         window_left=window_left,
+        sinks=sinks,
         o_sf_scale=o_sf_scale_float,
         out=output_trtllm,
     )
@@ -217,11 +239,13 @@ def test_flashinfer_trtllm_decode_with_baseline(
         output_trtllm = output_trtllm.reshape(-1, query.shape[1], query.shape[2])
 
     if q_quant_dtype == FP8_DTYPE and o_quant_dtype == FP4_DTYPE:
-        rtol, atol = 3e-1, 1e0
+        rtol, atol = 7e-2, 9e-2
     elif q_quant_dtype == FP8_DTYPE and o_quant_dtype == FP8_DTYPE:
-        rtol, atol = 5e-2, 7e-2
-    else:
+        rtol, atol = 2e-2, 4e-2
+    elif q_quant_dtype == FP8_DTYPE and o_quant_dtype == dtype:
         rtol, atol = 1e-2, 2e-2
+    else:
+        rtol, atol = 1e-2, 1e-2
 
     (
         torch.testing.assert_close(output, output_trtllm, atol=atol, rtol=rtol),
@@ -239,6 +263,7 @@ def test_flashinfer_trtllm_decode_with_baseline(
 @pytest.mark.parametrize("block_size", BLOCK_SIZE)
 @pytest.mark.parametrize("window_left", WINDOW_LEFT)
 @pytest.mark.parametrize("soft_cap", [None])
+@pytest.mark.parametrize("has_sinks", HAS_SINKS)
 @torch.inference_mode
 def test_flashinfer_trtllm_prefill_with_baseline(
     dtype: torch.dtype,
@@ -253,9 +278,10 @@ def test_flashinfer_trtllm_prefill_with_baseline(
     block_size: int,
     window_left: int,
     soft_cap: Optional[float],
+    has_sinks: bool,
 ) -> None:
     torch.set_default_device("cuda")
-    current_platform.seed_everything(0)
+    current_platform.seed_everything(42)
 
     q_quant_dtype, kv_quant_dtype, o_quant_dtype = quant_dtypes
     q_quant_dtype = q_quant_dtype or dtype
@@ -297,7 +323,7 @@ def test_flashinfer_trtllm_prefill_with_baseline(
         q_scale = 1.0
         ref_query = query
 
-    kv_lens = torch.randint(0, max_kv_len, (batch_size,), dtype=torch.int32)
+    kv_lens = torch.randint(1, max_kv_len, (batch_size,), dtype=torch.int32)
     kv_lens[-1] = max_kv_len
 
     seq_lens = kv_lens + q_lens
@@ -336,28 +362,36 @@ def test_flashinfer_trtllm_prefill_with_baseline(
     workspace_buffer = torch.zeros(128 * 1024 * 1024, dtype=torch.int8)
 
     # Baseline Prefill
+    if has_sinks:
+        sinks = torch.rand(num_qo_heads, dtype=torch.float32) * 5
+        wrapper = flashinfer.BatchAttentionWithAttentionSinkWrapper(
+            float_workspace_buffer=workspace_buffer, kv_layout=kv_layout, backend="fa2"
+        )
+    else:
+        sinks = None
         wrapper = flashinfer.BatchPrefillWithPagedKVCacheWrapper(
-        workspace_buffer, kv_layout
+            float_workspace_buffer=workspace_buffer, kv_layout=kv_layout, backend="fa2"
         )
+
     wrapper.plan(
-        q_indptr,
-        kv_indptr,
-        kv_indices,
-        kv_last_page_lens,
-        num_qo_heads,
-        num_kv_heads,
-        head_size,
-        block_size,
+        qo_indptr=q_indptr,
+        paged_kv_indptr=kv_indptr,
+        paged_kv_indices=kv_indices,
+        paged_kv_last_page_len=kv_last_page_lens,
+        num_qo_heads=num_qo_heads,
+        num_kv_heads=num_kv_heads,
+        head_dim_qk=head_size,
+        page_size=block_size,
         causal=True,
         sm_scale=sm_scale,
-        q_data_type=dtype,
-        kv_data_type=dtype,
         window_left=window_left,
         logits_soft_cap=soft_cap,
+        q_data_type=dtype,
+        kv_data_type=dtype,
     )
-
     output = torch.empty(ref_query.shape, dtype=dtype)
-    wrapper.run(ref_query, ref_kv_cache, out=output)
+    wrapper.run(ref_query, ref_kv_cache, sinks, sm_scale, out=output)
+
     o_scale = 1.0
     o_sf_scale_float = None
     if o_quant_dtype == FP8_DTYPE:
@@ -395,6 +429,7 @@ def test_flashinfer_trtllm_prefill_with_baseline(
         cum_seq_lens_q=q_indptr,
         cum_seq_lens_kv=kv_indptr,
         window_left=window_left,
+        sinks=sinks,
         o_sf_scale=o_sf_scale_float,
         out=output_trtllm,
     )
@@ -410,11 +445,11 @@ def test_flashinfer_trtllm_prefill_with_baseline(
         output_trtllm = output_trtllm.reshape(-1, query.shape[1], query.shape[2])
 
     if q_quant_dtype == FP8_DTYPE and o_quant_dtype == FP4_DTYPE:
-        rtol, atol = 4e-1, 1e0
+        rtol, atol = 1e-1, 2e-1
     elif q_quant_dtype == FP8_DTYPE and o_quant_dtype == FP8_DTYPE:
-        rtol, atol = 5e-2, 7e-2
-    elif q_quant_dtype == FP8_DTYPE and o_quant_dtype == dtype:
         rtol, atol = 4e-2, 6e-2
+    elif q_quant_dtype == FP8_DTYPE and o_quant_dtype == dtype:
+        rtol, atol = 2e-2, 3e-2
     else:
         rtol, atol = 1e-2, 1e-2
 

vllm/utils/flashinfer.py

@@ -269,11 +269,6 @@ def use_trtllm_attention(
 
     # Must use TRTLLM attention if query is FP8 quantized
     if q_dtype == current_platform.fp8_dtype():
-        if has_sinks:
-            raise RuntimeError(
-                "TRTLLM FP8-qkv kernel is not supported for attention sinks. "
-                "Use kv_cache_dtype=auto for now."
-            )
         logger.info_once("Using TRTLLM attention (query is quantized).")
         return True