RC specific changes (commandr compilation, CAR size increase, tags fetch from upstream, config defaults)

Author: micah-wil

docker/Dockerfile.rocm

@@ -27,9 +27,11 @@ FROM base AS fetch_vllm_1
 ARG VLLM_REPO="https://github.com/vllm-project/vllm.git"
 ARG VLLM_BRANCH="main"
 ONBUILD RUN git clone ${VLLM_REPO} \
-	    && cd vllm \
-	    && git fetch -v --prune -- origin ${VLLM_BRANCH} \
-	    && git checkout FETCH_HEAD
+        && cd vllm \
+        && git fetch -v --prune -- origin ${VLLM_BRANCH} \
+        && git checkout FETCH_HEAD \
+        && git remote add upstream "https://github.com/vllm-project/vllm.git" \
+        && git fetch upstream
 FROM fetch_vllm_${REMOTE_VLLM} AS fetch_vllm
 
 # -----------------------

vllm/config/compilation.py

@@ -75,11 +75,11 @@ class PassConfig:
     don't all have access to full configuration - that would create a cycle as
     the `PassManager` is set as a property of config."""
 
-    enable_fusion: bool = field(default_factory=lambda: not envs.VLLM_USE_V1)
+    enable_fusion: bool = field(default_factory=lambda: envs.VLLM_USE_V1)
     """Whether to enable the custom fusion (RMSNorm/SiluMul+quant) pass."""
     enable_attn_fusion: bool = False
     """Whether to enable the custom attention+quant fusion pass."""
-    enable_noop: bool = field(default_factory=lambda: not envs.VLLM_USE_V1)
+    enable_noop: bool = field(default_factory=lambda: envs.VLLM_USE_V1)
     """Whether to enable the custom no-op elimination pass."""
     enable_sequence_parallelism: bool = False
     """Whether to enable sequence parallelism."""
@@ -223,7 +223,7 @@ class CompilationConfig:
     constructor, e.g. `CompilationConfig(inductor_passes={"a": func})`."""
 
     # CudaGraph compilation
-    cudagraph_mode: Optional[CUDAGraphMode] = None
+    cudagraph_mode: Optional[CUDAGraphMode] = CUDAGraphMode.FULL
     """
     The mode of the cudagraph:
 
@@ -408,6 +408,16 @@ def __post_init__(self) -> None:
         count_all = self.custom_ops.count("all")
         assert count_none + count_all <= 1, "Can only specify 'none' or 'all'"
 
+        if "+rms_norm" not in self.custom_ops and \
+            "-rms_norm" not in self.custom_ops:
+            self.custom_ops.append("+rms_norm")
+        if "+silu_and_mul" not in self.custom_ops and \
+            "-silu_and_mul" not in self.custom_ops:
+            self.custom_ops.append("+silu_and_mul")
+        if "+quant_fp8" not in self.custom_ops and \
+            "-quant_fp8" not in self.custom_ops:
+            self.custom_ops.append("+quant_fp8")
+
         # TODO(zou3519/luka): There are 2 issues with auto-functionalization V2:
         # 1. A bug in PyTorch, fixed in 2.7:
         #    https://github.com/pytorch/pytorch/issues/147924

vllm/distributed/device_communicators/custom_all_reduce.py

@@ -54,7 +54,7 @@ class CustomAllreduce:
     def __init__(self,
                  group: ProcessGroup,
                  device: Union[int, str, torch.device],
-                 max_size=8192 * 1024) -> None:
+                 max_size=2 * 8192 * 1024) -> None:
         """
         Args:
             group: the process group to work on. If None, it will use the

vllm/envs.py

@@ -17,7 +17,7 @@
     VLLM_NCCL_SO_PATH: Optional[str] = None
     LD_LIBRARY_PATH: Optional[str] = None
     VLLM_USE_TRITON_FLASH_ATTN: bool = True
-    VLLM_V1_USE_PREFILL_DECODE_ATTENTION: bool = False
+    VLLM_V1_USE_PREFILL_DECODE_ATTENTION: bool = True
     VLLM_USE_AITER_UNIFIED_ATTENTION: bool = False
     VLLM_FLASH_ATTN_VERSION: Optional[int] = None
     LOCAL_RANK: int = 0
@@ -351,7 +351,7 @@ def get_vllm_port() -> Optional[int]:
     # the unified triton kernel.
     "VLLM_V1_USE_PREFILL_DECODE_ATTENTION":
     lambda:
-    (os.getenv("VLLM_V1_USE_PREFILL_DECODE_ATTENTION", "False").lower() in
+    (os.getenv("VLLM_V1_USE_PREFILL_DECODE_ATTENTION", "True").lower() in
      ("true", "1")),
 
     # Use AITER triton unified attention for V1 attention

vllm/model_executor/models/commandr.py

@@ -47,7 +47,6 @@
     row_parallel_weight_loader)
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.model_executor.utils import set_weight_attrs
-from vllm.platforms import current_platform
 from vllm.sequence import IntermediateTensors
 
 from .interfaces import SupportsLoRA, SupportsPP, SupportsQuant
@@ -57,7 +56,6 @@
                     maybe_prefix)
 
 
-@torch.compile(backend=current_platform.simple_compile_backend)
 def layer_norm_func(hidden_states, weight, variance_epsilon):
     input_dtype = hidden_states.dtype
     hidden_states = hidden_states.to(torch.float32)