vllm-cpu-amxbf16 0.9.1__cp312-cp312-manylinux_2_17_x86_64.whl

vllm/v1/attention/backends/pallas.py

@@ -0,0 +1,240 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from dataclasses import dataclass
+from typing import Any, Optional
+
+import torch
+# Required to register custom ops.
+import torch_xla.experimental.custom_kernel  # noqa: F401
+
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionLayer, AttentionType)
+from vllm.attention.backends.utils import CommonAttentionState
+from vllm.config import VllmConfig
+from vllm.logger import init_logger
+from vllm.utils import cdiv, next_power_of_2
+
+logger = init_logger(__name__)
+
+
+class PallasAttentionBackend(AttentionBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "PALLAS_VLLM_V1"
+
+    @staticmethod
+    def get_impl_cls() -> type["PallasAttentionBackendImpl"]:
+        return PallasAttentionBackendImpl
+
+    @staticmethod
+    def get_metadata_cls() -> type["PallasMetadata"]:
+        return PallasMetadata
+
+    @staticmethod
+    def get_state_cls() -> type["CommonAttentionState"]:
+        return CommonAttentionState
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> tuple[int, ...]:
+        return (num_blocks, block_size, num_kv_heads * 2, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: torch.Tensor,
+    ) -> None:
+        raise RuntimeError("swap_blocks is not used for the TPU backend.")
+
+    # In recent TPU generations, up to v6e, the SMEM size is 1MB. The
+    # block_tables within the PallasMetadata constitute almost the entire SMEM
+    # requirement. Its size is max_num_seqs * num_page_per_seq * 4 (Int). Here
+    # we simply make sure that the size is smaller than half of SMEM capacity.
+    @staticmethod
+    def get_min_page_size(vllm_config: VllmConfig) -> int:
+        max_num_page_per_req = (1024 * 1024 // 2 //
+                                vllm_config.scheduler_config.max_num_seqs // 4)
+        min_page_size = cdiv(vllm_config.model_config.max_model_len,
+                             max_num_page_per_req)
+        min_page_size = 1 << (min_page_size - 1).bit_length()
+        return min_page_size
+
+    # TPU has limited SREGs (scalar registers), if page_size is too small, we
+    # can spill SREGs easily which leads to bad performance. The strategy we
+    # apply here is trying to split max-model-len to 16 pages which make the
+    # spill less likely. Meanwhile we make sure the page size is in [16, 256].
+    @staticmethod
+    def get_page_size(vllm_config: VllmConfig) -> int:
+        page_size = next_power_of_2(
+            vllm_config.model_config.max_model_len) // 16
+        if page_size <= 16:
+            return 16
+        if page_size >= 256:
+            return 256
+        return page_size
+
+
+@dataclass
+class PallasMetadata:
+    # NOTE(sang): Definition of context_len, query_len, and seq_len.
+    # |---------- N-1 iteration --------|
+    # |---------------- N iteration ---------------------|
+    # |- tokenA -|......................|-- newTokens ---|
+    # |---------- context_len ----------|
+    # |-------------------- seq_len ---------------------|
+    #                                   |-- query_len ---|
+
+    # Used in the PallasAttentionBackendImpl
+    slot_mapping: torch.Tensor
+    block_tables: torch.Tensor
+    context_lens: torch.Tensor
+    query_start_loc: torch.Tensor
+    num_seqs: torch.Tensor
+
+
+class PallasAttentionBackendImpl(AttentionImpl):
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: Optional[list[float]],
+        sliding_window: Optional[int],
+        kv_cache_dtype: str,
+        blocksparse_params: Optional[dict[str, Any]] = None,
+        logits_soft_cap: Optional[float] = None,
+        attn_type: str = AttentionType.DECODER,
+        kv_sharing_target_layer_name: Optional[int] = None,
+        use_irope: bool = False,
+    ) -> None:
+        if use_irope:
+            logger.warning_once(
+                "Using irope in Pallas is not supported yet, it will fall back "
+                "to global attention for long context.")
+        if blocksparse_params is not None:
+            raise ValueError("Paged attention Pallas kernel does "
+                             "not support block-sparse attention.")
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.num_kv_heads = num_kv_heads
+        self.sliding_window = sliding_window
+        self.logits_soft_cap = logits_soft_cap
+        self.kv_sharing_target_layer_name = kv_sharing_target_layer_name
+
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+        if head_size % 128 != 0:
+            raise NotImplementedError("Head size must be a multiple of 128.")
+        if alibi_slopes is not None:
+            raise NotImplementedError("Alibi slopes is not supported.")
+        if kv_cache_dtype != "auto":
+            raise NotImplementedError("FP8 KV cache dtype is not supported.")
+        if blocksparse_params is not None:
+            raise NotImplementedError("Blocksparse is not supported.")
+
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError("Encoder self-attention and "
+                                      "encoder/decoder cross-attention "
+                                      "are not implemented for "
+                                      "PallasAttentionBackendImpl")
+
+        tpu_version = torch_xla.tpu.version()
+        if tpu_version < 4:
+            raise NotImplementedError("TPU version must be 4 or higher.")
+
+    def forward(
+        self,
+        layer: AttentionLayer,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: PallasMetadata,
+        output: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Forward pass with Pallas attention.
+
+        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 = [num_blocks, block_size, num_kv_heads * 2, head_size]
+            attn_metadata: Metadata for attention.
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+        # For determine_available_memory case.
+        if kv_cache.numel() == 0:
+            if output is None:
+                output = torch.ones_like(query)
+            return output
+
+        assert layer._k_scale_float == 1.0 and layer._v_scale_float == 1.0
+        num_tokens, hidden_size = query.shape
+        query = query.view(num_tokens, self.num_heads, self.head_size)
+
+        if self.kv_sharing_target_layer_name is None and kv_cache.numel() > 0:
+            # Write input keys and values to the KV cache.
+            # Skip this if sharing KV cache with an earlier attention layer.
+            slot_mapping = attn_metadata.slot_mapping
+            write_to_kv_cache(key, value, kv_cache, slot_mapping)
+
+        output = torch.ops.xla.ragged_paged_attention(
+            query,
+            kv_cache,
+            attn_metadata.context_lens,
+            attn_metadata.block_tables,
+            attn_metadata.query_start_loc,
+            attn_metadata.num_seqs,
+            # By default, the system utilizes optimized block size and
+            # vmem_limit_bytes parameters from the kernel repository. However,
+            # these can be manually adjusted for debugging if necessary.
+            num_kv_pages_per_block=None,
+            num_queries_per_block=None,
+            vmem_limit_bytes=None,
+            use_kernel=True,
+            sm_scale=self.scale,
+            sliding_window=self.sliding_window,
+            soft_cap=self.logits_soft_cap,
+        )
+
+        return output.reshape(num_tokens, hidden_size)
+
+
+def write_to_kv_cache(
+    key: torch.Tensor,
+    value: torch.Tensor,
+    kv_cache: torch.Tensor,
+    slot_mapping: torch.Tensor,
+) -> None:
+    """ Write the key and values to the KV cache.
+
+    Args:
+        key: shape = [num_tokens, num_kv_heads * head_size]
+        value: shape = [num_tokens, num_kv_heads *  head_size]
+        kv_cache = [num_blocks, block_size, num_kv_heads * 2, head_size]
+
+    """
+    _, _, num_combined_kv_heads, head_size = kv_cache.shape
+    num_kv_heads = num_combined_kv_heads // 2
+
+    key = key.view(-1, num_kv_heads, head_size)
+    value = value.view(-1, num_kv_heads, head_size)
+
+    kv = torch.cat([key, value], axis=-1).reshape(-1, num_combined_kv_heads,
+                                                  head_size)
+
+    torch.ops.xla.dynamo_set_buffer_donor_(kv_cache, True)
+
+    kv_cache = kv_cache.flatten(0, 1)
+    kv_cache.index_copy_(0, slot_mapping, kv)

vllm/v1/attention/backends/triton_attn.py

@@ -0,0 +1,285 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Attention layer with PagedAttention and Triton prefix prefill."""
+from typing import TYPE_CHECKING, Any, Optional
+
+import torch
+
+from vllm import _custom_ops as ops
+from vllm import envs
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionMetadata, AttentionType)
+from vllm.attention.ops.chunked_prefill_paged_decode import (
+    chunked_prefill_paged_decode)
+from vllm.attention.ops.paged_attn import PagedAttention
+from vllm.attention.ops.triton_unified_attention import unified_attention
+from vllm.logger import init_logger
+from vllm.platforms import current_platform
+from vllm.v1.attention.backends.flash_attn import (
+    FlashAttentionMetadata, FlashAttentionMetadataBuilder)
+from vllm.v1.kv_cache_interface import AttentionSpec
+from vllm.v1.worker.block_table import BlockTable
+
+if TYPE_CHECKING:
+    from vllm.v1.worker.gpu_model_runner import GPUModelRunner
+
+logger = init_logger(__name__)
+
+
+class TritonAttentionMetadataBuilder(FlashAttentionMetadataBuilder):
+
+    def __init__(self, runner: "GPUModelRunner", kv_cache_spec: AttentionSpec,
+                 block_table: BlockTable):
+        super().__init__(runner, kv_cache_spec, block_table)
+        self.aot_schedule = False
+
+
+class TritonAttentionBackend(AttentionBackend):
+
+    accept_output_buffer: bool = True
+
+    @staticmethod
+    def get_supported_head_sizes() -> list[int]:
+        return [32, 64, 96, 128, 160, 192, 224, 256]
+
+    @staticmethod
+    def get_name() -> str:
+        return "TRITON_ATTN_VLLM_V1"
+
+    @staticmethod
+    def get_impl_cls() -> type["TritonAttentionImpl"]:
+        return TritonAttentionImpl
+
+    @staticmethod
+    def get_metadata_cls() -> type["AttentionMetadata"]:
+        return FlashAttentionMetadata
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> tuple[int, ...]:
+        if block_size % 16 != 0:
+            raise ValueError("Block size must be a multiple of 16.")
+        return (2, num_blocks, block_size, num_kv_heads, head_size)
+
+    @staticmethod
+    def use_cascade_attention(*args, **kwargs) -> bool:
+        return False
+
+    @staticmethod
+    def get_builder_cls() -> type["TritonAttentionMetadataBuilder"]:
+        return TritonAttentionMetadataBuilder
+
+
+class TritonAttentionImpl(AttentionImpl):
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: Optional[list[float]],
+        sliding_window: Optional[int],
+        kv_cache_dtype: str,
+        blocksparse_params: Optional[dict[str, Any]] = None,
+        logits_soft_cap: Optional[float] = None,
+        attn_type: AttentionType = AttentionType.DECODER,
+        kv_sharing_target_layer_name: Optional[int] = None,
+        use_irope: bool = False,
+    ) -> None:
+        if blocksparse_params is not None:
+            raise ValueError(
+                "TritonAttention does not support block-sparse attention.")
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.num_kv_heads = num_kv_heads
+        if alibi_slopes is not None:
+            alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
+        self.alibi_slopes = alibi_slopes
+        if sliding_window is None:
+            self.sliding_window = (-1, -1)
+        else:
+            self.sliding_window = (sliding_window - 1, 0)
+        self.kv_cache_dtype = kv_cache_dtype
+        if logits_soft_cap is None:
+            # In flash-attn, setting logits_soft_cap as 0 means no soft cap.
+            logits_soft_cap = 0
+        self.logits_soft_cap = logits_soft_cap
+        self.kv_sharing_target_layer_name = kv_sharing_target_layer_name
+
+        self.use_irope = use_irope
+
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+
+        support_head_sizes = TritonAttentionBackend.get_supported_head_sizes()
+        if head_size not in support_head_sizes:
+            raise ValueError(
+                f"Head size {head_size} is not supported by TritonAttention. "
+                f"Supported head sizes are: {support_head_sizes}.")
+
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError("Encoder self-attention and "
+                                      "encoder/decoder cross-attention "
+                                      "are not implemented for "
+                                      "TritonAttentionImpl")
+
+        self.fp8_dtype = current_platform.fp8_dtype()
+        self.force_prefill_decode_attn = \
+            envs.VLLM_V1_USE_PREFILL_DECODE_ATTENTION
+
+    def forward(
+        self,
+        layer: torch.nn.Module,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: FlashAttentionMetadata,
+        output: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Forward pass with FlashAttention.
+
+        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]
+        """
+        assert output is not None, "Output tensor must be provided."
+
+        if attn_metadata is None:
+            # Profiling run.
+            return output
+
+        assert attn_metadata.use_cascade is False
+
+        # IMPORTANT!
+        # NOTE(woosuk): With piece-wise CUDA graphs, this method is executed in
+        # eager-mode PyTorch. Thus, we need to be careful about any CPU overhead
+        # in this method. For example, `view` and `slice` (or `[:n]`) operations
+        # are surprisingly slow even in the case they do not invoke any GPU ops.
+        # Minimize the PyTorch ops in this method as much as possible.
+        # Whenever making a change in this method, please benchmark the
+        # performance to make sure it does not introduce any overhead.
+
+        use_prefill_decode_attn = self.force_prefill_decode_attn
+        num_actual_tokens = attn_metadata.num_actual_tokens
+
+        if use_prefill_decode_attn:
+            key_cache, value_cache = PagedAttention.split_kv_cache(
+                kv_cache, self.num_kv_heads, self.head_size)
+        else:
+            key_cache, value_cache = kv_cache.unbind(0)
+
+        if self.kv_sharing_target_layer_name is None:
+            # Reshape the input keys and values and store them in the cache.
+            # Skip this if sharing KV cache with an earlier attention layer.
+            if use_prefill_decode_attn:
+                PagedAttention.write_to_paged_cache(
+                    key,
+                    value,
+                    key_cache,
+                    value_cache,
+                    attn_metadata.slot_mapping,
+                    self.kv_cache_dtype,
+                    layer._k_scale,
+                    layer._v_scale,
+                )
+            else:
+                torch.ops._C_cache_ops.reshape_and_cache_flash(
+                    key,
+                    value,
+                    key_cache,
+                    value_cache,
+                    attn_metadata.slot_mapping,
+                    self.kv_cache_dtype,
+                    layer._k_scale,
+                    layer._v_scale,
+                )
+
+        if self.kv_cache_dtype.startswith("fp8"):
+            key_cache = key_cache.view(self.fp8_dtype)
+            value_cache = value_cache.view(self.fp8_dtype)
+            num_tokens, num_heads, head_size = query.shape
+            assert layer._q_scale == 1.0, \
+                "A non 1.0 q_scale is not currently supported."
+            if not current_platform.is_rocm():
+                # Skip Q quantization on ROCm, since dequantizing back to
+                # f32 in the attention kernel is not supported.
+                query, _ = ops.scaled_fp8_quant(
+                    query.reshape(
+                        (num_tokens, num_heads * head_size)).contiguous(),
+                    layer._q_scale)
+            query = query.reshape((num_tokens, num_heads, head_size))
+
+        use_local_attn = \
+            (self.use_irope and attn_metadata.local_attn_metadata is not None)
+
+        if use_local_attn:
+            assert attn_metadata.local_attn_metadata is not None
+            local_metadata = attn_metadata.local_attn_metadata
+            cu_seqlens_q = local_metadata.local_query_start_loc
+            seqused_k = local_metadata.local_seqused_k
+            max_seqlen_q = local_metadata.local_max_query_len
+            max_seqlen_k = local_metadata.local_max_seq_len
+            block_table = local_metadata.local_block_table
+        else:
+            cu_seqlens_q = attn_metadata.query_start_loc
+            seqused_k = attn_metadata.seq_lens
+            max_seqlen_q = attn_metadata.max_query_len
+            max_seqlen_k = attn_metadata.max_seq_len
+            block_table = attn_metadata.block_table
+
+        if use_prefill_decode_attn:
+            # Compute attention and update output up to `num_actual_tokens`.
+            chunked_prefill_paged_decode(query=query[:num_actual_tokens],
+                                         key=key[:num_actual_tokens],
+                                         value=value[:num_actual_tokens],
+                                         output=output[:num_actual_tokens],
+                                         kv_cache_dtype=self.kv_cache_dtype,
+                                         key_cache=key_cache,
+                                         value_cache=value_cache,
+                                         block_table=block_table,
+                                         query_start_loc=cu_seqlens_q,
+                                         seq_lens=seqused_k,
+                                         max_seq_len=max_seqlen_k,
+                                         max_query_len=max_seqlen_q,
+                                         k_scale=layer._k_scale,
+                                         v_scale=layer._v_scale,
+                                         alibi_slopes=self.alibi_slopes,
+                                         sliding_window=self.sliding_window[0],
+                                         sm_scale=self.scale)
+
+        else:
+            descale_shape = (cu_seqlens_q.shape[0] - 1, key.shape[1])
+
+            unified_attention(
+                q=query[:num_actual_tokens],
+                k=key_cache,
+                v=value_cache,
+                out=output[:num_actual_tokens],
+                cu_seqlens_q=cu_seqlens_q,
+                max_seqlen_q=max_seqlen_q,
+                seqused_k=seqused_k,
+                max_seqlen_k=max_seqlen_k,
+                softmax_scale=self.scale,
+                causal=True,
+                alibi_slopes=self.alibi_slopes,
+                window_size=self.sliding_window,
+                block_table=block_table,
+                softcap=self.logits_soft_cap,
+                q_descale=None,  # Not supported
+                k_descale=layer._k_scale.expand(descale_shape),
+                v_descale=layer._v_scale.expand(descale_shape),
+            )
+
+        return output

vllm/v1/attention/backends/utils.py

@@ -0,0 +1,52 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from dataclasses import dataclass
+
+import torch
+
+
+@dataclass
+class CommonAttentionMetadata:
+    """
+    Attention metadata attributes that can be shared by layers in different KV
+    cache groups and thus having different block table.
+    """
+
+    query_start_loc: torch.Tensor
+    """(batch_size + 1,), the start location of each request in query Tensor"""
+    seq_lens: torch.Tensor
+    """(batch_size,), the length of each request including both computed tokens
+    and newly scheduled tokens"""
+
+
+def validate_kv_sharing_target(current_layer_name, target_layer_name,
+                               static_forward_context):
+    error_msg = (f"Specified KV sharing target layer for {current_layer_name} "
+                 f"is not valid: target layer {target_layer_name} ")
+
+    if current_layer_name == target_layer_name:
+        raise ValueError(error_msg +
+                         "cannot be the same as the current layer.")
+
+    if target_layer_name not in static_forward_context:
+        from vllm.model_executor.models.utils import extract_layer_index
+
+        # If target layer name is not in the static fwd context, it means either
+        # a) the target layer does not come BEFORE the current layer, or
+        # b) the target layer is not an Attention layer that exists in the model
+        current_layer_idx = extract_layer_index(current_layer_name)
+        target_layer_idx = extract_layer_index(target_layer_name)
+        if current_layer_idx <= target_layer_idx:
+            raise ValueError(error_msg + "must come before the current layer.")
+        else:
+            raise ValueError(error_msg +
+                             "is not a valid Attention layer in the model.")
+
+    # Currently KV sharing is only supported between layers of the same type
+    target_layer_attn_type = static_forward_context[
+        target_layer_name].attn_type
+    expected = static_forward_context[current_layer_name].attn_type
+    if target_layer_attn_type != expected:
+        raise ValueError(
+            error_msg +
+            f"must be the same type as the current layer ({expected}).")