[XPU] Prefer MXFP8 MoE XPU backend; forward softcap/ALIBI

docs/getting_started/installation/gpu.xpu.inc.md

@@ -12,6 +12,28 @@ vLLM initially supports basic model inference and serving on Intel GPU platform.
 !!! warning
     The provided vllm-xpu-kernels whl is Python3.12 specific so this version is a MUST.
 
+!!! note "XPU kernels runtime notes"
+    Install `vllm_xpu_kernels` from `requirements/xpu.txt` (release wheel by
+    default). Rebuild or reinstall that package when you change kernel configs.
+
+    Runtime notes (enforced by the kernels package):
+
+    - **Attention fail-closed (default):** missing FA2 / paged-decode AOT
+      shapes raise instead of silently falling back to a slow PyTorch
+      reference path. For eager debug only:
+      `export VLLM_XPU_ATTN_ALLOW_FALLBACK=1` (ignored under XPUGraph
+      capture). Softcap and ALiBi are not implemented on XPU FA2 and raise
+      when requested.
+    - **Page size 128:** default decode presets may include pagesize=128
+      shapes. Serve with `--block-size 128` to use them; otherwise keep the
+      usual page-64 path.
+    - **Native MXFP8 / block-FP8 MoE (Xe2):** fused MoE uses the native
+      grouped-GEMM path by default. Escape hatches:
+      `VLLM_XPU_FUSED_MOE_NATIVE_MXFP8=0`,
+      `VLLM_XPU_FUSED_MOE_NATIVE_BLOCK_FP8=0`, or
+      `VLLM_XPU_FUSED_MOE_USE_REF=1` (force the Python reference expert
+      loop for A/B).
+
 --8<-- [end:requirements]
 --8<-- [start:set-up-using-python]
 

vllm/_xpu_ops.py

@@ -994,6 +994,8 @@ def flash_attn_varlen_func(
             assert len(window_size) == 2
             real_window_size = (window_size[0], window_size[1])  # noqa: F841
 
+        # Forward softcap/alibi so the kernels package can fail closed when
+        # those features are requested but not implemented on XPU.
         return flash_attn_varlen_func(
             out=out,
             q=q,
@@ -1009,8 +1011,8 @@ def flash_attn_varlen_func(
             block_table=block_table,
             s_aux=s_aux,
             window_size=real_window_size,
-            # alibi_slopes = alibi_slopes,
-            # softcap=softcap,
+            alibi_slopes=alibi_slopes,
+            softcap=softcap if softcap is not None else 0.0,
             return_softmax_lse=return_softmax_lse,
             q_descale=q_descale,
             k_descale=k_descale,

vllm/envs.py

@@ -297,6 +297,9 @@
     VLLM_NIXL_EP_MAX_NUM_RANKS: int = 32
     VLLM_XPU_ENABLE_XPU_GRAPH: bool = False
     VLLM_XPU_USE_SAMPLER_KERNEL: bool = True
+    # Opt into slow PyTorch attention when an XPU FA2 shape is missing.
+    # Default (unset/0) is fail-closed in vllm-xpu-kernels.
+    VLLM_XPU_ATTN_ALLOW_FALLBACK: bool = False
     VLLM_LORA_ENABLE_DUAL_STREAM: bool = False
     VLLM_GPU_NIC_PCIE_MAPPING: str = ""
     VLLM_NIC_SELECTION_VARS: str = ""
@@ -2044,6 +2047,12 @@ def _resolve_rust_cli_path() -> str | None:
     "VLLM_XPU_USE_SAMPLER_KERNEL": lambda: bool(
         int(os.getenv("VLLM_XPU_USE_SAMPLER_KERNEL", "1"))
     ),
+    # Opt into PyTorch reference attention when an XPU FA2 AOT shape is
+    # missing. Default is fail-closed (raise). Eager debug only; ignored
+    # under XPUGraph capture. Read by vllm-xpu-kernels at runtime.
+    "VLLM_XPU_ATTN_ALLOW_FALLBACK": lambda: bool(
+        int(os.getenv("VLLM_XPU_ATTN_ALLOW_FALLBACK", "0"))
+    ),
     # Enable simple KV offload.
     "VLLM_USE_SIMPLE_KV_OFFLOAD": lambda: bool(
         int(os.getenv("VLLM_USE_SIMPLE_KV_OFFLOAD", "0"))

vllm/model_executor/layers/fused_moe/oracle/mxfp8.py

@@ -128,8 +128,15 @@ def select_mxfp8_moe_backend(
         )
         return backend, _select_kernel_cls(backend, config)
 
+    # Prefer XPU on Intel GPU before trying CUDA/ROCm backends.
+    backends = _SUPPORTED_BACKENDS
+    if current_platform.is_xpu():
+        backends = (Fp8MoeBackend.XPU,) + tuple(
+            b for b in _SUPPORTED_BACKENDS if b != Fp8MoeBackend.XPU
+        )
+
     # Auto-select: pick the first supported backend.
-    for backend in _SUPPORTED_BACKENDS:
+    for backend in backends:
         try:
             experts_cls = _select_kernel_cls(backend, config)
         except ValueError: