[ROCm] Add AITER RoPE + KV cache fusion for MLA prefill and decode

vllm/envs.py

@@ -51,9 +51,10 @@
     VLLM_CPU_OMP_THREADS_BIND: str = "auto"
     VLLM_CPU_NUM_OF_RESERVED_CPU: int | None = None
     VLLM_CPU_SGL_KERNEL: bool = False
-    VLLM_ZENTORCH_WEIGHT_PREPACK: bool = True
     VLLM_XLA_CACHE_PATH: str = os.path.join(VLLM_CACHE_ROOT, "xla_cache")
     VLLM_XLA_CHECK_RECOMPILATION: bool = False
+    VLLM_FUSED_MOE_CHUNK_SIZE: int = 16 * 1024
+    VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING: bool = True
     VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE: Literal["auto", "nccl", "shm"] = "auto"
     VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM: bool = False
     VLLM_USE_RAY_WRAPPED_PP_COMM: bool = True
@@ -97,7 +98,6 @@
     VLLM_ALLOW_RUNTIME_LORA_UPDATING: bool = False
     VLLM_SKIP_P2P_CHECK: bool = False
     VLLM_DISABLED_KERNELS: list[str] = []
-    VLLM_ENABLE_FLA_PACKED_RECURRENT_DECODE: bool = True
     VLLM_DISABLE_PYNCCL: bool = False
     VLLM_USE_OINK_OPS: bool = False
     VLLM_ROCM_USE_AITER: bool = False
@@ -117,6 +117,8 @@
     VLLM_ROCM_USE_SKINNY_GEMM: bool = True
     VLLM_ROCM_FP8_PADDING: bool = True
     VLLM_ROCM_MOE_PADDING: bool = True
+    VLLM_USE_AITER_FUSED: bool = True
+    VLLM_ROCM_CUSTOM_PAGED_ATTN: bool = True
     VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT: bool = False
     VLLM_ENABLE_V1_MULTIPROCESSING: bool = True
     VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
@@ -169,7 +171,7 @@
     VLLM_FLASHINFER_MOE_BACKEND: Literal["throughput", "latency", "masked_gemm"] = (
         "latency"
     )
-    VLLM_FLASHINFER_ALLREDUCE_BACKEND: Literal["auto", "trtllm", "mnnvl"] = "auto"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: Literal["auto", "trtllm", "mnnvl"] = "trtllm"
     VLLM_FLASHINFER_WORKSPACE_BUFFER_SIZE: int = 394 * 1024 * 1024
     VLLM_XGRAMMAR_CACHE_MB: int = 0
     VLLM_MSGPACK_ZERO_COPY_THRESHOLD: int = 256
@@ -246,8 +248,6 @@
     VLLM_ELASTIC_EP_SCALE_UP_LAUNCH: bool = False
     VLLM_ELASTIC_EP_DRAIN_REQUESTS: bool = False
     VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS: bool = False
-    VLLM_NIXL_EP_MAX_NUM_RANKS: int = 32
-    VLLM_XPU_ENABLE_XPU_GRAPH: bool = False
 
 
 def get_default_cache_root():
@@ -295,16 +295,6 @@ def use_aot_compile() -> bool:
     )
 
 
-def use_mega_aot_artifact():
-    from vllm.utils.torch_utils import is_torch_equal_or_newer
-
-    default_value = (
-        "1" if is_torch_equal_or_newer("2.12.0.dev") and use_aot_compile() else "0"
-    )
-
-    return os.environ.get("VLLM_USE_MEGA_AOT_ARTIFACT", default_value) == "1"
-
-
 def env_with_choices(
     env_name: str,
     default: str | None,
@@ -628,7 +618,10 @@ def _get_or_set_default() -> str:
     # Enable loading compiled models directly from cached standalone compile artifacts
     # without re-splitting graph modules. This reduces overhead during model
     # loading by using reconstruct_serializable_fn_from_mega_artifact.
-    "VLLM_USE_MEGA_AOT_ARTIFACT": use_mega_aot_artifact,
+    "VLLM_USE_MEGA_AOT_ARTIFACT": lambda: os.environ.get(
+        "VLLM_USE_MEGA_AOT_ARTIFACT", "0"
+    )
+    == "1",
     # local rank of the process in the distributed setting, used to determine
     # the GPU device id
     "LOCAL_RANK": lambda: int(os.environ.get("LOCAL_RANK", "0")),
@@ -719,11 +712,6 @@ def _get_or_set_default() -> str:
     else None,
     # (CPU backend only) whether to use SGL kernels, optimized for small batch.
     "VLLM_CPU_SGL_KERNEL": lambda: bool(int(os.getenv("VLLM_CPU_SGL_KERNEL", "0"))),
-    # (Zen CPU backend) eagerly prepack weights into ZenDNN blocked layout
-    # at model load time. Eliminates per-inference layout conversion overhead.
-    "VLLM_ZENTORCH_WEIGHT_PREPACK": lambda: bool(
-        int(os.getenv("VLLM_ZENTORCH_WEIGHT_PREPACK", "1"))
-    ),
     # If the env var is set, Ray Compiled Graph uses the specified
     # channel type to communicate between workers belonging to
     # different pipeline-parallel stages.
@@ -841,6 +829,15 @@ def _get_or_set_default() -> str:
     ),
     # Enable SPMD mode for TPU backend.
     "VLLM_XLA_USE_SPMD": lambda: bool(int(os.getenv("VLLM_XLA_USE_SPMD", "0"))),
+    "VLLM_FUSED_MOE_CHUNK_SIZE": lambda: int(
+        os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", str(16 * 1024))
+    ),
+    # Control whether to use fused MoE activation chunking. Current chunking
+    # logic is incompatible with torch.compile and causes IMA. See issue
+    # https://github.com/vllm-project/vllm/issues/19631.
+    "VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING": lambda: bool(
+        int(os.getenv("VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING", "1"))
+    ),
     # If set, the OpenAI API server will stay alive even after the underlying
     # AsyncLLMEngine errors and stops serving requests
     "VLLM_KEEP_ALIVE_ON_ENGINE_DEATH": lambda: bool(
@@ -910,9 +907,6 @@ def _get_or_set_default() -> str:
     "VLLM_DISABLED_KERNELS": lambda: []
     if "VLLM_DISABLED_KERNELS" not in os.environ
     else os.environ["VLLM_DISABLED_KERNELS"].split(","),
-    "VLLM_ENABLE_FLA_PACKED_RECURRENT_DECODE": lambda: bool(
-        int(os.getenv("VLLM_ENABLE_FLA_PACKED_RECURRENT_DECODE", "1"))
-    ),
     # Disable pynccl (using torch.distributed instead)
     "VLLM_DISABLE_PYNCCL": lambda: (
         os.getenv("VLLM_DISABLE_PYNCCL", "False").lower() in ("true", "1")
@@ -993,6 +987,14 @@ def _get_or_set_default() -> str:
     "VLLM_ROCM_USE_AITER_TRITON_GEMM": lambda: (
         os.getenv("VLLM_ROCM_USE_AITER_TRITON_GEMM", "True").lower() in ("true", "1")
     ),
+    # Enable AITER fused decode kernel for MLA (ROCm only, decode path only)
+    # Enable AITER fused kernels for MLA (ROCm only, prefill and decode)
+    # Fuses: RoPE + concat + KV cache write (prefill) or BMM + RoPE +
+    # concat + KV cache write (decode) in ONE kernel
+    # By default is enabled for AMD GPUs with FP8 support.
+    "VLLM_USE_AITER_FUSED": lambda: (
+        os.getenv("VLLM_USE_AITER_FUSED", "True").lower() in ("true", "1")
+    ),
     # use rocm skinny gemms
     "VLLM_ROCM_USE_SKINNY_GEMM": lambda: (
         os.getenv("VLLM_ROCM_USE_SKINNY_GEMM", "True").lower() in ("true", "1")
@@ -1001,6 +1003,10 @@ def _get_or_set_default() -> str:
     "VLLM_ROCM_FP8_PADDING": lambda: bool(int(os.getenv("VLLM_ROCM_FP8_PADDING", "1"))),
     # Pad the weights for the moe kernel
     "VLLM_ROCM_MOE_PADDING": lambda: bool(int(os.getenv("VLLM_ROCM_MOE_PADDING", "1"))),
+    # custom paged attention kernel for MI3* cards
+    "VLLM_ROCM_CUSTOM_PAGED_ATTN": lambda: (
+        os.getenv("VLLM_ROCM_CUSTOM_PAGED_ATTN", "True").lower() in ("true", "1")
+    ),
     # Whether to use the shuffled kv cache layout
     "VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT": lambda: (
         os.getenv("VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT", "False").lower() in ("true", "1")
@@ -1305,9 +1311,14 @@ def _get_or_set_default() -> str:
         ["throughput", "latency", "masked_gemm"],
     ),
     # Flashinfer fused allreduce backend.
+    # "auto" will default to "mnnvl", which performs mostly same/better than "trtllm".
+    # But "mnnvl" backend does not support fuse with quantization.
+    # TODO: Default is "trtllm" right now because "mnnvl" has issues with cudagraph:
+    # https://github.com/vllm-project/vllm/issues/35772
+    # Should switch back to "auto" if the issue is resolved.
     "VLLM_FLASHINFER_ALLREDUCE_BACKEND": env_with_choices(
         "VLLM_FLASHINFER_ALLREDUCE_BACKEND",
-        "auto",
+        "trtllm",
         ["auto", "trtllm", "mnnvl"],
     ),
     # Control the workspace buffer size for the FlashInfer backend.
@@ -1640,14 +1651,6 @@ def _get_or_set_default() -> str:
     "VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS": lambda: bool(
         int(os.getenv("VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS", "0"))
     ),
-    # NIXL EP environment variables
-    "VLLM_NIXL_EP_MAX_NUM_RANKS": lambda: int(
-        os.getenv("VLLM_NIXL_EP_MAX_NUM_RANKS", "32")
-    ),
-    # Whether enable XPU graph on Intel GPU
-    "VLLM_XPU_ENABLE_XPU_GRAPH": lambda: bool(
-        int(os.getenv("VLLM_XPU_ENABLE_XPU_GRAPH", "0"))
-    ),
 }
 
 
@@ -1784,7 +1787,6 @@ def compile_factors() -> dict[str, object]:
         "VLLM_V1_OUTPUT_PROC_CHUNK_SIZE",
         "VLLM_CPU_KVCACHE_SPACE",
         "VLLM_CPU_MOE_PREPACK",
-        "VLLM_ZENTORCH_WEIGHT_PREPACK",
         "VLLM_TEST_FORCE_LOAD_FORMAT",
         "VLLM_ENABLE_CUDA_COMPATIBILITY",
         "VLLM_CUDA_COMPATIBILITY_PATH",