[Core][Kernels] Enable FP8 KV Cache with Flashinfer backend. + BugFix for kv_cache_dtype=auto (#7985)

Co-authored-by: Simon Mo <simon.mo@hey.com>
Co-authored-by: Cody Yu <hao.yu.cody@gmail.com>

tests/kernels/test_flashinfer.py

@@ -73,11 +73,14 @@ def ref_paged_attn(
 @pytest.mark.parametrize("dtype", DTYPES)
 @pytest.mark.parametrize("soft_cap", [None, 30.0, 50.0])
 @torch.inference_mode
-def test_flashinfer_decode_with_paged_kv(kv_lens: List[int],
+def test_flashinfer_decode_with_paged_kv(
-                                         num_heads: Tuple[int,
+    kv_lens: List[int],
-                                                          int], head_size: int,
+    num_heads: Tuple[int, int],
-                                         dtype: torch.dtype, block_size: int,
+    head_size: int,
-                                         soft_cap: Optional[float]) -> None:
+    dtype: torch.dtype,
+    block_size: int,
+    soft_cap: Optional[float],
+) -> None:
     torch.set_default_device("cuda")
     torch.cuda.manual_seed_all(0)
     num_seqs = len(kv_lens)
@@ -88,6 +91,7 @@ def test_flashinfer_decode_with_paged_kv(kv_lens: List[int],
     scale = head_size**-0.5
 
     query = torch.randn(num_seqs, num_query_heads, head_size, dtype=dtype)
+
     key_value_cache = torch.randn(NUM_BLOCKS,
                                   2,
                                   block_size,
@@ -125,7 +129,7 @@ def test_flashinfer_decode_with_paged_kv(kv_lens: List[int],
     wrapper = flashinfer.\
         BatchDecodeWithPagedKVCacheWrapper(workspace_buffer, "NHD",
                 use_tensor_cores=(
-                    (num_query_heads//num_kv_heads) not in (1, 2, 4, 8))
+                    (num_query_heads//num_kv_heads) > 4)
                 )
     wrapper.begin_forward(kv_indptr,
                           kv_indices,
@@ -249,3 +253,215 @@ def test_flashinfer_prefill_with_paged_kv(seq_lens: List[Tuple[int, int]],
                                 soft_cap=soft_cap)
     torch.testing.assert_close(output, ref_output, atol=1e-2, rtol=1e-2), \
         f"{torch.max(torch.abs(output - ref_output))}"
+
+
+@pytest.mark.parametrize("seq_lens", [[(1, 132), (5, 18)]])
+@pytest.mark.parametrize("num_heads", [(32, 8), (6, 1)])
+@pytest.mark.parametrize("head_size", HEAD_SIZES)
+@pytest.mark.parametrize("block_size", BLOCK_SIZES)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("soft_cap", [None, 30.0, 50.0])
+def test_flashinfer_prefill_with_paged_fp8_kv(
+        seq_lens: List[Tuple[int, int]], num_heads: Tuple[int, int],
+        head_size: int, dtype: torch.dtype, block_size: int,
+        soft_cap: Optional[float]) -> None:
+    torch.set_default_device("cuda")
+    torch.cuda.manual_seed_all(0)
+    num_seqs = len(seq_lens)
+    query_lens = [x[0] for x in seq_lens]
+    kv_lens = [x[1] for x in seq_lens]
+    num_query_heads = num_heads[0]
+    num_kv_heads = num_heads[1]
+    assert num_query_heads % num_kv_heads == 0
+    max_kv_len = max(kv_lens)
+    scale = head_size**-0.5
+
+    kv_cache_dtype = torch.float8_e4m3fn
+
+    query = torch.randn(sum(query_lens),
+                        num_query_heads,
+                        head_size,
+                        dtype=dtype)
+    NUM_BLOCKS_FP8 = 2048
+    key_value_cache = torch.randn(NUM_BLOCKS_FP8,
+                                  2,
+                                  block_size,
+                                  num_kv_heads,
+                                  head_size,
+                                  dtype=dtype)
+    key_cache, value_cache = torch.chunk(key_value_cache, 2, dim=1)
+    key_cache /= head_size**0.5
+    value_cache /= head_size**0.5
+
+    k_scale = key_cache.amax().item() / 448.0
+    v_scale = value_cache.amax().item() / 448.0
+
+    kv_cache_fp8 = torch.cat([key_cache / k_scale, value_cache / v_scale],
+                             dim=1).to(kv_cache_dtype)
+
+    assert (kv_cache_fp8.shape == key_value_cache.shape)
+    max_num_blocks_per_seq = (max_kv_len + block_size - 1) // block_size
+    block_tables = torch.randint(0,
+                                 NUM_BLOCKS_FP8,
+                                 (num_seqs, max_num_blocks_per_seq),
+                                 dtype=torch.int32)
+
+    qo_indptr = [0]
+    kv_indptr = [0]
+    kv_indices = []
+    kv_last_page_lens = []
+    for i in range(num_seqs):
+        seq_len = kv_lens[i]
+        assert seq_len > 0
+        num_blocks = (seq_len + block_size - 1) // block_size
+        kv_indices.extend(block_tables[i, :num_blocks])
+        kv_indptr.append(kv_indptr[-1] + num_blocks)
+        kv_last_page_len = seq_len % block_size
+        if kv_last_page_len == 0:
+            kv_last_page_len = block_size
+        kv_last_page_lens.append(kv_last_page_len)
+        qo_indptr.append(qo_indptr[-1] + query_lens[i])
+
+    qo_indptr = torch.tensor(qo_indptr, dtype=torch.int32)
+    kv_indptr = torch.tensor(kv_indptr, dtype=torch.int32)
+    kv_indices = torch.tensor(kv_indices, dtype=torch.int32)
+    kv_last_page_lens = torch.tensor(kv_last_page_lens, dtype=torch.int32)
+
+    workspace_buffer = torch.empty(128 * 1024 * 1024, dtype=torch.int8)
+    wrapper = flashinfer.BatchPrefillWithPagedKVCacheWrapper(
+        workspace_buffer, "NHD")
+    wrapper.begin_forward(
+        qo_indptr,
+        kv_indptr,
+        kv_indices,
+        kv_last_page_lens,
+        num_query_heads,
+        num_kv_heads,
+        head_size,
+        block_size,
+    )
+
+    output = wrapper.forward(query,
+                             kv_cache_fp8,
+                             logits_soft_cap=soft_cap,
+                             k_scale=k_scale,
+                             v_scale=v_scale)
+
+    ref_output = ref_paged_attn(query=query,
+                                key_cache=key_cache.squeeze(1),
+                                value_cache=value_cache.squeeze(1),
+                                query_lens=query_lens,
+                                kv_lens=kv_lens,
+                                block_tables=block_tables,
+                                scale=scale,
+                                soft_cap=soft_cap)
+    del query
+    del block_tables
+    # verify prefill fp8
+    torch.testing.assert_close(output, ref_output, atol=1e-2, rtol=1e-2), \
+        f"{torch.max(torch.abs(output - ref_output))}"
+
+
+@pytest.mark.parametrize("kv_lens", [[1328, 18, 463], [1, 54, 293, 70]])
+@pytest.mark.parametrize("num_heads", [(32, 8), (64, 8), (6, 1)])
+@pytest.mark.parametrize("head_size", HEAD_SIZES)
+@pytest.mark.parametrize("block_size", BLOCK_SIZES)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("soft_cap", [None, 30.0, 50.0])
+@torch.inference_mode
+def test_flashinfer_decode_with_paged_fp8_kv(
+    kv_lens: List[int],
+    num_heads: Tuple[int, int],
+    head_size: int,
+    dtype: torch.dtype,
+    block_size: int,
+    soft_cap: Optional[float],
+) -> None:
+    # test doesn't work for num_heads = (16,16)
+    torch.set_default_device("cuda")
+    torch.cuda.manual_seed_all(0)
+    num_seqs = len(kv_lens)
+    num_query_heads = num_heads[0]
+    num_kv_heads = num_heads[1]
+    assert num_query_heads % num_kv_heads == 0
+    max_kv_len = max(kv_lens)
+    scale = head_size**-0.5
+    use_tensor_cores = (num_query_heads // num_kv_heads) > 4
+    kv_cache_dtype = torch.float8_e4m3fn
+
+    query = torch.randn(num_seqs, num_query_heads, head_size, dtype=dtype)
+    NUM_BLOCKS_FP8 = 2048
+    key_value_cache = torch.randn(NUM_BLOCKS_FP8,
+                                  2,
+                                  block_size,
+                                  num_kv_heads,
+                                  head_size,
+                                  dtype=dtype)
+    key_cache, value_cache = torch.chunk(key_value_cache, 2, dim=1)
+    key_cache /= head_size**0.5
+    value_cache /= head_size**0.5
+
+    k_scale = key_cache.amax().item() / 448.0
+    v_scale = value_cache.amax().item() / 448.0
+
+    key_cache_fp8 = (key_cache / k_scale).to(kv_cache_dtype)
+    value_cache_fp8 = (value_cache / v_scale).to(kv_cache_dtype)
+    assert (key_cache_fp8.shape[1] == 1 and value_cache_fp8.shape[1] == 1)
+    kv_cache_fp8 = torch.cat([key_cache_fp8, value_cache_fp8], dim=1)
+
+    max_num_blocks_per_seq = (max_kv_len + block_size - 1) // block_size
+    block_tables = torch.randint(0,
+                                 NUM_BLOCKS_FP8,
+                                 (num_seqs, max_num_blocks_per_seq),
+                                 dtype=torch.int32)
+
+    kv_indptr = [0]
+    kv_indices = []
+    kv_last_page_lens = []
+    for i in range(num_seqs):
+        seq_len = kv_lens[i]
+        assert seq_len > 0
+        num_blocks = (seq_len + block_size - 1) // block_size
+        kv_indices.extend(block_tables[i, :num_blocks])
+        kv_indptr.append(kv_indptr[-1] + num_blocks)
+        kv_last_page_len = seq_len % block_size
+        if kv_last_page_len == 0:
+            kv_last_page_len = block_size
+        kv_last_page_lens.append(kv_last_page_len)
+
+    kv_indptr = torch.tensor(kv_indptr, dtype=torch.int32)
+    kv_indices = torch.tensor(kv_indices, dtype=torch.int32)
+    kv_last_page_lens = torch.tensor(kv_last_page_lens, dtype=torch.int32)
+
+    workspace_buffer = torch.empty(128 * 1024 * 1024, dtype=torch.int8)
+    wrapper = flashinfer.\
+        BatchDecodeWithPagedKVCacheWrapper(workspace_buffer, "NHD",
+                    use_tensor_cores=use_tensor_cores)
+    wrapper.begin_forward(kv_indptr,
+                          kv_indices,
+                          kv_last_page_lens,
+                          num_query_heads,
+                          num_kv_heads,
+                          head_size,
+                          block_size,
+                          "NONE",
+                          data_type=dtype)
+    output = wrapper.forward(query,
+                             kv_cache_fp8,
+                             logits_soft_cap=soft_cap,
+                             k_scale=k_scale,
+                             v_scale=v_scale)
+    key_cache = key_value_cache[:, 0, :, :, :].squeeze(1)
+    value_cache = key_value_cache[:, 1, :, :, :].squeeze(1)
+
+    ref_output = ref_paged_attn(query=query,
+                                key_cache=key_cache,
+                                value_cache=value_cache,
+                                query_lens=[1] * num_seqs,
+                                kv_lens=kv_lens,
+                                block_tables=block_tables,
+                                scale=scale,
+                                soft_cap=soft_cap)
+    # Temporary fix: Increasing the tolerance. Seems like a flashinfer issue
+    torch.testing.assert_close(output, ref_output, atol=2e-2, rtol=1e-2), \
+        f"{torch.max(torch.abs(output - ref_output))}"

vllm/attention/backends/flashinfer.py

@@ -83,6 +83,15 @@ class FlashInferBackend(AttentionBackend):
     def get_supported_head_sizes() -> List[int]:
         return [64, 128, 256]
 
+    @staticmethod
+    def get_fp8_dtype_for_flashinfer(kv_cache_dtype: str) -> torch.dtype:
+        if kv_cache_dtype in ("fp8", "fp8_e4m3"):
+            return torch.float8_e4m3fn
+        elif kv_cache_dtype == "fp8_e5m2":
+            return torch.float8_e5m2
+        else:
+            raise ValueError(f"Unrecognized FP8 dtype: {kv_cache_dtype}")
+
 
 class FlashInferState(AttentionState):
 
@@ -177,9 +186,9 @@ class FlashInferState(AttentionState):
             self._graph_decode_workspace_buffer, _indptr_buffer,
             self._graph_indices_buffer, _last_page_len_buffer, "NHD",
             use_tensor_cores)
-        kv_cache_dtype = get_kv_cache_torch_dtype(
-            self.runner.kv_cache_dtype, self.runner.model_config.dtype)
 
+        kv_cache_dtype = FlashInferBackend.get_fp8_dtype_for_flashinfer(
+            self.runner.kv_cache_dtype)
         paged_kv_indptr_tensor_host = torch.arange(0,
                                                    batch_size + 1,
                                                    dtype=torch.int32)
@@ -340,7 +349,7 @@ class FlashInferMetadata(AttentionMetadata):
                 self.page_size,
                 # Disable flashinfer's pos encoding and use vllm's rope.
                 pos_encoding_mode="NONE",
-                data_type=self.data_type)
+            )
 
     def asdict_zerocopy(self,
                         skip_fields: Optional[Set[str]] = None
@@ -366,7 +375,8 @@ class FlashInferMetadata(AttentionMetadata):
     def decode_metadata(self) -> Optional["FlashInferMetadata"]:
         # Currently chunked prefill is not supported
         if self.num_prefills > 0:
-            assert self.num_decode_tokens == 0
+            assert self.num_decode_tokens == 0, (
+                "Chunked prefill is not supported with flashinfer yet.")
             return None
 
         return self
@@ -578,6 +588,7 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
 
         kv_cache_dtype = get_kv_cache_torch_dtype(
             self.runner.kv_cache_dtype, self.runner.model_config.dtype)
+
         return FlashInferMetadata(
             num_prefills=self.num_prefills,
             slot_mapping=slot_mapping_tensor,
@@ -661,7 +672,6 @@ class FlashInferImpl(AttentionImpl):
         if attn_metadata.num_decode_tokens > 0:
             assert attn_metadata.num_prefill_tokens == 0, (
                 "Chunked prefill is not supported with flashinfer yet.")
-
         if kv_cache is not None:
             # Use the same reshape and cache kernel as flash attention.
             ops.reshape_and_cache_flash(
@@ -674,6 +684,12 @@ class FlashInferImpl(AttentionImpl):
                 k_scale,
                 v_scale,
             )
+            # The FlashInfer api requires data to be in fp8_e4m3 or fp8_e5m2
+            # to process the cache when the kv_cache_dtype is fp8
+            if self.kv_cache_dtype.startswith("fp8"):
+                torch_dtype = FlashInferBackend.get_fp8_dtype_for_flashinfer(
+                    self.kv_cache_dtype)
+                kv_cache = kv_cache.view(torch_dtype)
 
         query = query.contiguous(
         )  # Flashinfer requires query to be contiguous
@@ -711,5 +727,7 @@ class FlashInferImpl(AttentionImpl):
                 query,
                 kv_cache,
                 sm_scale=self.scale,
-                logits_soft_cap=self.logits_soft_cap)
+                logits_soft_cap=self.logits_soft_cap,
+                k_scale=k_scale,
+                v_scale=v_scale)
         return output.view(num_tokens, hidden_size)

vllm/attention/selector.py

@@ -226,6 +226,10 @@ def which_attn_to_use(
         elif kv_cache_dtype is not None and kv_cache_dtype.startswith("fp8"):
             logger.info(
                 "Cannot use FlashAttention-2 backend for FP8 KV cache.")
+            logger.warning(
+                "Please use FlashInfer backend with FP8 KV Cache for "
+                "better performance by setting environment variable  "
+                "VLLM_ATTENTION_BACKEND=FLASHINFER")
             selected_backend = _Backend.XFORMERS
         elif block_size % 16 != 0:
             logger.info(