[XPU] support Triton Attention backend on Intel GPU (vllm-project#24149)

Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>

Author: jikunshang

.buildkite/scripts/hardware_ci/run-xpu-test.sh

@@ -30,10 +30,11 @@ docker run \
     bash -c '
     set -e
     echo $ZE_AFFINITY_MASK
-    VLLM_USE_V1=1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
-    VLLM_USE_V1=1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 -O3 -O.cudagraph_mode=NONE
-    VLLM_USE_V1=1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend ray
-    VLLM_USE_V1=1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 -O3 -O.cudagraph_mode=NONE
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend ray
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
+    VLLM_ATTENTION_BACKEND=TRITON_ATTN_VLLM_V1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
     cd tests
     pytest -v -s v1/core
     pytest -v -s v1/engine

vllm/_ipex_ops.py

@@ -242,10 +242,9 @@ def reshape_and_cache_flash(
         k_scale_float: float = 1.0,
         v_scale_float: float = 1.0,
     ) -> None:
-        assert kv_cache_dtype == "auto"
-        # TODO: support FP8 kv cache.
         ipex.llm.modules.PagedAttention.reshape_and_cache_flash(
-            key, value, key_cache, value_cache, slot_mapping)
+            key, value, key_cache, value_cache, slot_mapping, kv_cache_dtype,
+            k_scale_float, v_scale_float)
 
     @staticmethod
     def flash_attn_varlen_func(

vllm/attention/ops/paged_attn.py

@@ -6,9 +6,14 @@
 
 import torch
 
-from vllm import _custom_ops as ops
+from vllm.platforms import current_platform
 from vllm.triton_utils import HAS_TRITON
 
+if current_platform.is_cuda_alike():
+    from vllm import _custom_ops as ops
+elif current_platform.is_xpu():
+    from vllm._ipex_ops import ipex_ops as ops
+
 if HAS_TRITON:
     from vllm.attention.ops.prefix_prefill import context_attention_fwd
 

vllm/platforms/xpu.py

@@ -37,14 +37,38 @@ def get_attn_backend_cls(cls, selected_backend: _Backend, head_size: int,
                              dtype: torch.dtype, kv_cache_dtype: Optional[str],
                              block_size: int, use_v1: bool, use_mla: bool,
                              has_sink: bool) -> str:
-        if selected_backend is not None and selected_backend != _Backend.IPEX:
-            logger.info("Cannot use %s backend on XPU.", selected_backend)
         use_v1 = envs.VLLM_USE_V1
         if not use_v1:
             raise ValueError("XPU backend only supports V1.")
+        TRITON_ATTN_VLLM_V1 = "vllm.v1.attention.backends.triton_attn.TritonAttentionBackend"  # noqa: E501
+        FLASH_ATTN_V1 = "vllm.v1.attention.backends.flash_attn.FlashAttentionBackend"  # noqa: E501
+        if selected_backend == _Backend.TRITON_ATTN_VLLM_V1:
+            logger.info_once("Using Triton backend on V1 engine.")
+            return TRITON_ATTN_VLLM_V1
+        elif selected_backend == _Backend.FLASH_ATTN:
+            logger.info_once("Using Flash Attention backend on V1 engine.")
+            return FLASH_ATTN_V1
+        elif selected_backend:
+            raise ValueError(
+                f"Invalid attention backend for {cls.device_name}, "
+                f"with use_v1: {use_v1} use_mla: {use_mla}")
+
         logger.info("Using Flash Attention backend on V1 engine.")
         return "vllm.v1.attention.backends.flash_attn.FlashAttentionBackend"
 
+    @classmethod
+    def is_kv_cache_dtype_supported(cls, kv_cache_dtype: str,
+                                    model_config: "ModelConfig") -> bool:
+        """
+        Check if the kv_cache_dtype is supported.
+        XPU only support fp8 kv cache with triton backend.
+        """
+        if envs.is_set("VLLM_ATTENTION_BACKEND") and \
+            envs.VLLM_ATTENTION_BACKEND == "TRITON_ATTN_VLLM_V1":
+            return kv_cache_dtype in ["fp8_e4m3", "fp8_e5m2", "fp8"]
+
+        return False
+
     @classmethod
     def set_device(cls, device: torch.device) -> None:
         """

vllm/v1/attention/backends/triton_attn.py

@@ -7,7 +7,6 @@
 
 import torch
 
-from vllm import _custom_ops as ops
 from vllm import envs
 from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
                                               AttentionMetadata, AttentionType)
@@ -23,6 +22,11 @@
                                               CommonAttentionMetadata)
 from vllm.v1.kv_cache_interface import AttentionSpec
 
+if current_platform.is_cuda_alike():
+    from vllm import _custom_ops as ops
+elif current_platform.is_xpu():
+    from vllm._ipex_ops import ipex_ops as ops
+
 logger = init_logger(__name__)
 
 
@@ -337,7 +341,7 @@ def forward(
                     layer._v_scale,
                 )
             else:
-                torch.ops._C_cache_ops.reshape_and_cache_flash(
+                ops.reshape_and_cache_flash(
                     key,
                     value,
                     key_cache,
@@ -354,9 +358,10 @@ def forward(
             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.
+            if current_platform.is_cuda():
+                # Skip Q quantization on ROCm and XPU, enable this on cuda
+                # only, 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(),