vllm-npu 0.4.2__py3-none-any.whl

vllm/__init__.py

@@ -0,0 +1,23 @@
+"""vLLM: a high-throughput and memory-efficient inference engine for LLMs"""
+import vllm_npu
+from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
+from vllm.engine.async_llm_engine import AsyncLLMEngine
+from vllm.engine.llm_engine import LLMEngine
+from vllm.entrypoints.llm import LLM
+from vllm.executor.ray_utils import initialize_ray_cluster
+from vllm.model_executor.models import ModelRegistry
+from vllm.outputs import CompletionOutput, RequestOutput
+from vllm.sampling_params import SamplingParams
+__version__ = "0.4.2"
+__all__ = [
+    "LLM",
+    "ModelRegistry",
+    "SamplingParams",
+    "RequestOutput",
+    "CompletionOutput",
+    "LLMEngine",
+    "EngineArgs",
+    "AsyncLLMEngine",
+    "AsyncEngineArgs",
+    "initialize_ray_cluster",
+]

vllm/_custom_ops.py

@@ -0,0 +1,251 @@
+from typing import Dict, Optional, Tuple
+
+import torch
+
+try:
+    from vllm._C import cache_ops as vllm_cache_ops
+    from vllm._C import ops as vllm_ops
+except ImportError:
+    pass
+
+
+# activation ops
+def silu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
+    vllm_ops.silu_and_mul(out, x)
+
+
+def gelu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
+    vllm_ops.gelu_and_mul(out, x)
+
+
+def gelu_tanh_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
+    vllm_ops.gelu_tanh_and_mul(out, x)
+
+
+def gelu_fast(out: torch.Tensor, x: torch.Tensor) -> None:
+    vllm_ops.gelu_fast(out, x)
+
+
+def gelu_new(out: torch.Tensor, x: torch.Tensor) -> None:
+    vllm_ops.gelu_new(out, x)
+
+
+# page attention ops
+def paged_attention_v1(
+    out: torch.Tensor,
+    query: torch.Tensor,
+    key_cache: torch.Tensor,
+    value_cache: torch.Tensor,
+    num_kv_heads: int,
+    scale: float,
+    block_tables: torch.Tensor,
+    seq_lens: torch.Tensor,
+    block_size: int,
+    max_seq_len: int,
+    alibi_slopes: Optional[torch.Tensor],
+    kv_cache_dtype: str,
+    kv_scale: float,
+) -> None:
+    vllm_ops.paged_attention_v1(out, query, key_cache, value_cache,
+                                num_kv_heads, scale, block_tables, seq_lens,
+                                block_size, max_seq_len, alibi_slopes,
+                                kv_cache_dtype, kv_scale)
+
+
+def paged_attention_v2(
+    out: torch.Tensor,
+    exp_sum: torch.Tensor,
+    max_logits: torch.Tensor,
+    tmp_out: torch.Tensor,
+    query: torch.Tensor,
+    key_cache: torch.Tensor,
+    value_cache: torch.Tensor,
+    num_kv_heads: int,
+    scale: float,
+    block_tables: torch.Tensor,
+    seq_lens: torch.Tensor,
+    block_size: int,
+    max_seq_len: int,
+    alibi_slopes: Optional[torch.Tensor],
+    kv_cache_dtype: str,
+    kv_scale: float,
+) -> None:
+    vllm_ops.paged_attention_v2(out, exp_sum, max_logits, tmp_out, query,
+                                key_cache, value_cache, num_kv_heads, scale,
+                                block_tables, seq_lens, block_size,
+                                max_seq_len, alibi_slopes, kv_cache_dtype,
+                                kv_scale)
+
+
+# pos encoding ops
+def rotary_embedding(
+    positions: torch.Tensor,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    head_size: int,
+    cos_sin_cache: torch.Tensor,
+    is_neox: bool,
+) -> None:
+    vllm_ops.rotary_embedding(positions, query, key, head_size, cos_sin_cache,
+                              is_neox)
+
+
+def batched_rotary_embedding(positions: torch.Tensor, query: torch.Tensor,
+                             key: torch.Tensor, head_size: int,
+                             cos_sin_cache: torch.Tensor, is_neox: bool,
+                             rot_dim: int,
+                             cos_sin_cache_offsets: torch.Tensor) -> None:
+    vllm_ops.batched_rotary_embedding(positions, query, key, head_size,
+                                      cos_sin_cache, is_neox, rot_dim,
+                                      cos_sin_cache_offsets)
+
+
+# layer norm ops
+def rms_norm(out: torch.Tensor, input: torch.Tensor, weight: torch.Tensor,
+             epsilon: float) -> None:
+    vllm_ops.rms_norm(out, input, weight, epsilon)
+
+
+def fused_add_rms_norm(input: torch.Tensor, residual: torch.Tensor,
+                       weight: torch.Tensor, epsilon: float) -> None:
+    vllm_ops.fused_add_rms_norm(input, residual, weight, epsilon)
+
+
+# quantization ops
+# awq
+def awq_dequantize(qweight: torch.Tensor, scales: torch.Tensor,
+                   zeros: torch.Tensor, split_k_iters: int, thx: int,
+                   thy: int) -> torch.Tensor:
+    return vllm_ops.awq_dequantize(qweight, scales, zeros, split_k_iters, thx,
+                                   thy)
+
+
+def awq_gemm(input: torch.Tensor, qweight: torch.Tensor, qzeros: torch.Tensor,
+             scales: torch.Tensor, split_k_iters: int) -> torch.Tensor:
+    return vllm_ops.awq_gemm(input, qweight, qzeros, scales, split_k_iters)
+
+
+# gptq
+def gptq_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
+              b_gptq_qzeros: torch.Tensor, b_gptq_scales: torch.Tensor,
+              b_g_idx: torch.Tensor, use_exllama: bool,
+              bit: int) -> torch.Tensor:
+    return vllm_ops.gptq_gemm(a, b_q_weight, b_gptq_qzeros, b_gptq_scales,
+                              b_g_idx, use_exllama, bit)
+
+
+def gptq_shuffle(q_weight: torch.Tensor, q_perm: torch.Tensor,
+                 bit: int) -> None:
+    vllm_ops.gptq_shuffle(q_weight, q_perm, bit)
+
+
+# squeezellm
+def squeezellm_gemm(vec: torch.Tensor, mat: torch.Tensor, mul: torch.Tensor,
+                    lookup_table: torch.Tensor) -> None:
+    vllm_ops.squeezellm_gemm(vec, mat, mul, lookup_table)
+
+
+# marlin
+def marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
+                b_scales: torch.Tensor, workspace: torch.Tensor, size_m: int,
+                size_n: int, size_k: int) -> torch.Tensor:
+    return vllm_ops.marlin_gemm(a, b_q_weight, b_scales, workspace, size_m,
+                                size_n, size_k)
+
+
+# aqlm
+def aqlm_gemm(input: torch.Tensor, codes: torch.Tensor,
+              codebooks: torch.Tensor, scales: torch.Tensor,
+              codebook_partition_sizes: torch.Tensor,
+              bias: Optional[torch.Tensor]) -> torch.Tensor:
+    return vllm_ops.aqlm_gemm(input, codes, codebooks, scales,
+                              codebook_partition_sizes, bias)
+
+
+def aqlm_dequant(codes: torch.Tensor, codebooks: torch.Tensor,
+                 codebook_partition_sizes: torch.Tensor) -> torch.Tensor:
+    return vllm_ops.aqlm_dequant(codes, codebooks, codebook_partition_sizes)
+
+
+# gptq_marlin
+def gptq_marlin_repack(b_q_weight: torch.Tensor, perm: torch.Tensor,
+                       size_k: int, size_n: int,
+                       num_bits: int) -> torch.Tensor:
+    return vllm_ops.gptq_marlin_repack(b_q_weight, perm, size_k, size_n,
+                                       num_bits)
+
+
+def gptq_marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
+                     b_scales: torch.Tensor, g_idx: torch.Tensor,
+                     perm: torch.Tensor, workspace: torch.Tensor,
+                     num_bits: int, size_m: int, size_n: int, size_k: int,
+                     is_k_full: bool) -> torch.Tensor:
+    return vllm_ops.gptq_marlin_gemm(a, b_q_weight, b_scales, g_idx, perm,
+                                     workspace, num_bits, size_m, size_n,
+                                     size_k, is_k_full)
+
+
+# fp8
+def scaled_fp8_quant(
+    input: torch.Tensor,
+    scale: Optional[torch.Tensor] = None,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    output = torch.empty_like(input, dtype=torch.float8_e4m3fn)
+    if scale is None:
+        scale = torch.zeros(1, device=input.device, dtype=torch.float32)
+        vllm_ops.dynamic_scaled_fp8_quant(output, input, scale)
+    else:
+        vllm_ops.static_scaled_fp8_quant(output, input, scale)
+    return output, scale
+
+
+# moe
+def moe_align_block_size(topk_ids: torch.Tensor, num_experts: int,
+                         block_size: int, sorted_token_ids: torch.Tensor,
+                         experts_ids: torch.Tensor,
+                         num_tokens_post_pad: torch.Tensor) -> None:
+    vllm_ops.moe_align_block_size(topk_ids, num_experts, block_size,
+                                  sorted_token_ids, experts_ids,
+                                  num_tokens_post_pad)
+
+
+def reshape_and_cache(
+    key: torch.Tensor,
+    value: torch.Tensor,
+    key_cache: torch.Tensor,
+    value_cache: torch.Tensor,
+    slot_mapping: torch.Tensor,
+    kv_cache_dtype: str,
+    kv_scale: float,
+) -> None:
+    vllm_cache_ops.reshape_and_cache(key, value, key_cache, value_cache,
+                                     slot_mapping, kv_cache_dtype, kv_scale)
+
+
+def reshape_and_cache_flash(
+    key: torch.Tensor,
+    value: torch.Tensor,
+    key_cache: torch.Tensor,
+    value_cache: torch.Tensor,
+    slot_mapping: torch.Tensor,
+    kv_cache_dtype: str,
+) -> None:
+    vllm_cache_ops.reshape_and_cache_flash(key, value, key_cache, value_cache,
+                                           slot_mapping, kv_cache_dtype)
+
+
+def copy_blocks(key_caches: torch.Tensor, value_caches: torch.Tensor,
+                block_mapping: torch.Tensor) -> None:
+    vllm_cache_ops.copy_blocks(key_caches, value_caches, block_mapping)
+
+
+def swap_blocks(src: torch.Tensor, dst: torch.Tensor,
+                block_mapping: Dict[int, int]) -> None:
+    vllm_cache_ops.swap_blocks(src, dst, block_mapping)
+
+
+def convert_fp8(output: torch.Tensor, input: torch.Tensor) -> None:
+    vllm_cache_ops.convert_fp8(output, input)
+
+
+#TODO: cuda_utils, custom_ar

vllm/attention/__init__.py

@@ -0,0 +1,13 @@
+from vllm.attention.backends.abstract import (AttentionBackend,
+                                              AttentionMetadata,
+                                              AttentionMetadataPerStage)
+from vllm.attention.layer import Attention
+from vllm.attention.selector import get_attn_backend
+
+__all__ = [
+    "AttentionBackend",
+    "AttentionMetadata",
+    "Attention",
+    "get_attn_backend",
+    "AttentionMetadataPerStage",
+]

vllm/attention/backends/abstract.py

@@ -0,0 +1,127 @@
+from abc import ABC, abstractmethod
+from dataclasses import dataclass, fields
+from typing import (Any, Dict, Generic, List, Optional, Set, Tuple, Type,
+                    TypeVar)
+
+import torch
+
+
+class AttentionBackend(ABC):
+    """Abstract class for attention backends."""
+
+    @staticmethod
+    @abstractmethod
+    def get_impl_cls() -> Type["AttentionImpl"]:
+        raise NotImplementedError
+
+    @staticmethod
+    @abstractmethod
+    def make_metadata(*args, **kwargs) -> "AttentionMetadataPerStage":
+        raise NotImplementedError
+
+    @staticmethod
+    @abstractmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        raise NotImplementedError
+
+    @staticmethod
+    @abstractmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: Dict[int, int],
+    ) -> None:
+        raise NotImplementedError
+
+    @staticmethod
+    @abstractmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: Dict[int, List[int]],
+    ) -> None:
+        raise NotImplementedError
+
+
+@dataclass
+class AttentionMetadataPerStage:
+    """Attention metadata for a specific stage. I.e., prefill or decode."""
+
+    def asdict_zerocopy(self,
+                        skip_fields: Optional[Set[str]] = None
+                        ) -> Dict[str, Any]:
+        """Similar to dataclasses.asdict, but avoids deepcopying."""
+        if skip_fields is None:
+            skip_fields = set()
+        # Note that if we add dataclasses as fields, they will need
+        # similar handling.
+        return {
+            field.name: getattr(self, field.name)
+            for field in fields(self) if field.name not in skip_fields
+        }
+
+
+T = TypeVar("T", bound=AttentionMetadataPerStage)
+
+
+@dataclass
+class AttentionMetadata(Generic[T]):
+    """Attention metadata for prefill and decode batched together."""
+    # Total number of prefill requests.
+    num_prefills: int
+    # Number of prefill tokens.
+    num_prefill_tokens: int
+    # Number of decode tokens. Note that it is equivalent to the number of
+    # decode requests.
+    num_decode_tokens: int
+    # The attention metadata for prefill requests in a batch.
+    # None if there's no prefill requests in a batch.
+    prefill_metadata: Optional[T]
+    # The attention metadata for decode requests in a batch.
+    # None if there's no decode requests in a batch.
+    decode_metadata: Optional[T]
+    # (num_tokens,). The indices of the token slots that input tokens will be
+    # stored into. E.g., if `slot_mapping` is [35, 2, 17] and the block size
+    # is 16, the three tokens are stored in the 3rd slot in block 2, 2nd slot
+    # in block 0, and 1st slot in block 1, respectively.
+    slot_mapping: torch.Tensor
+    # The kv cache's data type.
+    kv_cache_dtype: str
+
+    def __post_init__(self):
+        if self.num_prefill_tokens > 0:
+            assert self.num_prefills > 0
+            assert self.prefill_metadata is not None
+        if self.num_decode_tokens > 0:
+            assert self.decode_metadata is not None
+
+
+class AttentionImpl(ABC):
+
+    @abstractmethod
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: Optional[int] = None,
+        alibi_slopes: Optional[List[float]] = None,
+        sliding_window: Optional[int] = None,
+    ) -> None:
+        raise NotImplementedError
+
+    @abstractmethod
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+        kv_scale: float,
+    ) -> torch.Tensor:
+        raise NotImplementedError

vllm/attention/backends/flash_attn.py

@@ -0,0 +1,271 @@
+"""Attention layer with Flash and PagedAttention.
+
+NOTE(woosuk): At the moment, this file includes a lot of duplicated code from
+XFormers backend. The duplicated code will be removed once we use flash-attn or
+flashinfer for all the attention operations.
+"""
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple, Type
+
+import torch
+from flash_attn import flash_attn_varlen_func
+
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionMetadata,
+                                              AttentionMetadataPerStage)
+from vllm.attention.ops.paged_attn import (PagedAttention,
+                                           PagedAttentionMetadata)
+
+
+class FlashAttentionBackend(AttentionBackend):
+
+    @staticmethod
+    def get_impl_cls() -> Type["FlashAttentionImpl"]:
+        return FlashAttentionImpl
+
+    @staticmethod
+    def make_metadata(*args, **kwargs) -> "FlashAttentionMetadata":
+        return FlashAttentionMetadata(*args, **kwargs)
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return PagedAttention.get_kv_cache_shape(num_blocks, block_size,
+                                                 num_kv_heads, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: Dict[int, int],
+    ) -> None:
+        PagedAttention.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: Dict[int, List[int]],
+    ) -> None:
+        PagedAttention.copy_blocks(kv_caches, src_to_dists)
+
+
+@dataclass
+class FlashAttentionMetadata(AttentionMetadataPerStage,
+                             PagedAttentionMetadata):
+    """Metadata for FlashAttentionBackend.
+
+    NOTE: Any python object stored here is not updated when it is
+    cuda-graph replayed. If you have values that need to be changed
+    dynamically, it should be stored in tensor. The tensor has to be
+    updated from `CUDAGraphRunner.forward` API.
+    """
+    # Currently, input sequences can only contain all prompts
+    # or all decoding. True if all sequences are prompts.
+    is_prompt: bool
+    # (batch_size,). The sequence length per sequence. Sequence length means
+    # the computed tokens + new tokens None if it is a decoding.
+    seq_lens: Optional[List[int]]
+    # seq_lens stored as a tensor.
+    seq_lens_tensor: Optional[torch.Tensor]
+
+    # NOTE(sang): Definition of context_len, query_len, and seq_len.
+    # |---------- N-1 iteration --------|
+    # |---------------- N iteration ---------------------|
+    # |- tokenA -|......................|-- newTokens ---|
+    # |---------- context_len ----------|
+    # |-------------------- seq_len ----------------------|
+    #                                   |-- query_len ---|
+
+    # Maximum query length in the batch.
+    max_query_len: Optional[int]
+    # Maximum sequence length in the batch.
+    max_seq_len: Optional[int]
+    # (batch_size + 1,). The cumulative subquery lengths of the sequences in
+    # the batch, used to index into subquery. E.g., if the subquery length
+    # is [4, 6], it is [0, 4, 10].
+    subquery_start_loc: Optional[torch.Tensor]
+    # (batch_size + 1,). The cumulative sequence lengths of the sequences in
+    # the batch, used to index into sequence. E.g., if the sequence length is
+    # [4, 6], it is [0, 4, 10].
+    seq_start_loc: Optional[torch.Tensor]
+    # (batch_size,) A tensor of context lengths (tokens that are computed
+    # so far).
+    context_lens_tensor: Optional[torch.Tensor]
+
+    # Whether or not if cuda graph is enabled.
+    # Cuda-graph is currently enabled for decoding only.
+    # TODO(woosuk): Move `use_cuda_graph` out since it's unrelated to attention.
+    use_cuda_graph: bool
+
+
+class FlashAttentionImpl(AttentionImpl):
+    """
+    If the input tensors contain prompt tokens, the layout is as follows:
+    |<--------------- num_prefill_tokens ----------------->|	
+    |<--prefill_0-->|<--prefill_1-->|...|<--prefill_N-1--->|
+
+    Otherwise, the layout is as follows:	
+    |<----------------- num_decode_tokens ------------------>|	
+    |<--decode_0-->|..........|<--decode_M-1-->|<--padding-->|
+
+    Generation tokens can contain padding when cuda-graph is used.
+    Currently, prompt tokens don't contain any padding.
+
+    The prompts might have different lengths, while the generation tokens
+    always have length 1.
+
+    If chunked prefill is enabled, prefill tokens and decode tokens can be
+    batched together in a flattened 1D query.
+
+    |<----- num_prefill_tokens ---->|<------- num_decode_tokens --------->|
+    |<-prefill_0->|...|<-prefill_N-1->|<--decode_0-->|...|<--decode_M-1-->|
+
+    Currently, cuda graph is disabled for chunked prefill, meaning there's no
+    padding between prefill and decode tokens.
+    """
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: Optional[int] = None,
+        alibi_slopes: Optional[List[float]] = None,
+        sliding_window: Optional[int] = None,
+    ) -> None:
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
+        self.sliding_window = ((sliding_window, sliding_window)
+                               if sliding_window is not None else (-1, -1))
+        if alibi_slopes is not None:
+            alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
+        self.alibi_slopes = alibi_slopes
+
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+
+        suppored_head_sizes = PagedAttention.get_supported_head_sizes()
+        if head_size not in suppored_head_sizes:
+            raise ValueError(
+                f"Head size {head_size} is not supported by PagedAttention. "
+                f"Supported head sizes are: {suppored_head_sizes}.")
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata[FlashAttentionMetadata],
+        kv_scale: float,
+    ) -> torch.Tensor:
+        """Forward pass with FlashAttention and PagedAttention.
+
+        Args:
+            query: shape = [num_tokens, num_heads * head_size]
+            key: shape = [num_tokens, num_kv_heads * head_size]
+            value: shape = [num_tokens, num_kv_heads * head_size]
+            kv_cache = [2, num_blocks, block_size * num_kv_heads * head_size]
+            attn_metadata: Metadata for attention.
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+        num_tokens, hidden_size = query.shape
+        # Reshape the query, key, and value tensors.
+        query = query.view(-1, self.num_heads, self.head_size)
+        key = key.view(-1, self.num_kv_heads, self.head_size)
+        value = value.view(-1, self.num_kv_heads, self.head_size)
+
+        if kv_cache is not None:
+            key_cache, value_cache = PagedAttention.split_kv_cache(
+                kv_cache, self.num_kv_heads, self.head_size)
+
+            # Reshape the input keys and values and store them in the cache.
+            # If kv_cache is not provided, the new key and value tensors are
+            # not cached. This happens during the initial memory profiling run.
+            PagedAttention.write_to_paged_cache(key, value, key_cache,
+                                                value_cache,
+                                                attn_metadata.slot_mapping,
+                                                attn_metadata.kv_cache_dtype,
+                                                kv_scale)
+
+        num_prefill_tokens = attn_metadata.num_prefill_tokens
+        num_decode_tokens = attn_metadata.num_decode_tokens
+        assert key.shape[0] == num_prefill_tokens + num_decode_tokens
+        assert value.shape[0] == num_prefill_tokens + num_decode_tokens
+
+        output = torch.empty_like(query)
+        # Query for decode. KV is not needed because it is already cached.
+        decode_query = query[num_prefill_tokens:]
+        # QKV for prefill.
+        query = query[:num_prefill_tokens]
+        key = key[:num_prefill_tokens]
+        value = value[:num_prefill_tokens]
+
+        assert query.shape[0] == num_prefill_tokens
+        assert decode_query.shape[0] == num_decode_tokens
+
+        if prefill_meta := attn_metadata.prefill_metadata:
+            # Prompt run.
+            if kv_cache is None or prefill_meta.block_tables.numel() == 0:
+                # normal attention
+                # When block_tables are not filled, it means q and k are the
+                # prompt, and they have the same length.
+                out = flash_attn_varlen_func(
+                    q=query,
+                    k=key,
+                    v=value,
+                    cu_seqlens_q=prefill_meta.seq_start_loc,
+                    cu_seqlens_k=prefill_meta.seq_start_loc,
+                    max_seqlen_q=prefill_meta.max_seq_len,
+                    max_seqlen_k=prefill_meta.max_seq_len,
+                    softmax_scale=self.scale,
+                    causal=True,
+                    window_size=self.sliding_window,
+                    alibi_slopes=self.alibi_slopes,
+                )
+                assert output[:num_prefill_tokens].shape == out.shape
+                output[:num_prefill_tokens] = out
+            else:
+                # prefix-enabled attention
+                # TODO(Hai) this triton kernel has regression issue (broke) to
+                # deal with different data types between KV and FP8 KV cache,
+                # to be addressed separately.
+                output[:num_prefill_tokens] = PagedAttention.forward_prefix(
+                    query,
+                    key,
+                    value,
+                    key_cache,
+                    value_cache,
+                    prefill_meta.block_tables,
+                    prefill_meta.subquery_start_loc,
+                    prefill_meta.seq_lens_tensor,
+                    prefill_meta.context_lens_tensor,
+                    prefill_meta.max_query_len,
+                    self.alibi_slopes,
+                    self.sliding_window[0],
+                )
+        if decode_meta := attn_metadata.decode_metadata:
+            # Decoding run.
+            output[num_prefill_tokens:] = PagedAttention.forward_decode(
+                decode_query,
+                key_cache,
+                value_cache,
+                decode_meta.block_tables,
+                decode_meta.seq_lens_tensor,
+                decode_meta.max_seq_len,
+                attn_metadata.kv_cache_dtype,
+                self.num_kv_heads,
+                self.scale,
+                self.alibi_slopes,
+                kv_scale,
+            )
+
+        # Reshape the output tensor.
+        return output.view(num_tokens, hidden_size)