cleanup

Author: gshtras

requirements/rocm.txt

@@ -5,7 +5,6 @@ numba == 0.60.0; python_version == '3.9' # v0.61 doesn't support Python 3.9. Req
 numba == 0.61.2; python_version > '3.9'
 
 # Dependencies for AMD GPUs
-numpy==1.26.4
 boto3
 botocore
 datasets

vllm/attention/backends/rocm_flash_attn.py

@@ -535,7 +535,7 @@ def __init__(
                 f"Head size {head_size} is not supported by PagedAttention. "
                 f"Supported head sizes are: {supported_head_sizes}.")
 
-        self.use_naive_attn = envs.VLLM_USE_SDPA_ATTENTION  # Default False
+        self.use_naive_attn = False
         # NOTE: Allow for switching between Triton and CK. Defaulting to triton.
         self.use_triton_flash_attn = envs.VLLM_USE_TRITON_FLASH_ATTN
         if self.use_triton_flash_attn:

vllm/envs.py

@@ -14,7 +14,6 @@
     VLLM_RINGBUFFER_WARNING_INTERVAL: int = 60
     VLLM_NCCL_SO_PATH: Optional[str] = None
     LD_LIBRARY_PATH: Optional[str] = None
-    VLLM_USE_SDPA_ATTENTION: bool = False
     VLLM_USE_TRITON_FLASH_ATTN: bool = True
     VLLM_USE_ROCM_CUSTOM_PAGED_ATTN_FP8_OUT: bool = True
     VLLM_USE_ROCM_FP8_FLASH_ATTN: bool = False
@@ -287,11 +286,6 @@ def get_vllm_port() -> Optional[int]:
     "LD_LIBRARY_PATH":
     lambda: os.environ.get("LD_LIBRARY_PATH", None),
 
-    # flag to control if vllm should use naive scaled dot-product attention
-    "VLLM_USE_SDPA_ATTENTION":
-    lambda: (os.environ.get("VLLM_USE_SDPA_ATTENTION", "False").lower() in
-             ("true", "1")),
-
     # flag to control if vllm should use triton flash attention
     "VLLM_USE_TRITON_FLASH_ATTN":
     lambda: (os.environ.get("VLLM_USE_TRITON_FLASH_ATTN", "True").lower() in