diff --git a/tests/kernels/attention/test_use_trtllm_attention.py b/tests/kernels/attention/test_use_trtllm_attention.py index a58a650fdd..fba18fe46e 100644 --- a/tests/kernels/attention/test_use_trtllm_attention.py +++ b/tests/kernels/attention/test_use_trtllm_attention.py @@ -72,7 +72,7 @@ def test_supports_sm100_with_artifactory(_art, _cap): @patch("vllm.envs.VLLM_BATCH_INVARIANT", False) @patch( - "vllm.utils.flashinfer.current_platform.is_device_capability", + "vllm.utils.flashinfer.current_platform.is_device_capability_family", return_value=False, ) def test_supports_non_sm100_platform(_cap): @@ -81,7 +81,7 @@ def test_supports_non_sm100_platform(_cap): @patch("vllm.envs.VLLM_BATCH_INVARIANT", False) @patch( - "vllm.utils.flashinfer.current_platform.is_device_capability", + "vllm.utils.flashinfer.current_platform.is_device_capability_family", return_value=True, ) @patch("vllm.utils.flashinfer.has_nvidia_artifactory", return_value=False) diff --git a/tools/pre_commit/generate_attention_backend_docs.py b/tools/pre_commit/generate_attention_backend_docs.py index 9e14f8739d..bbbf4f4b64 100644 --- a/tools/pre_commit/generate_attention_backend_docs.py +++ b/tools/pre_commit/generate_attention_backend_docs.py @@ -235,11 +235,10 @@ def _resolve_import_to_file( def _find_cc_in_function(tree: ast.AST, func_name: str) -> str | None: - """Find a compute capability from is_device_capability*() calls in a function. + """Find a compute capability from is_device_capability_family() calls in a function. - Handles two patterns: - - is_device_capability_family(N): "M.x" (e.g. 100 -> "10.x") - - is_device_capability(N): "M.m" (e.g. 100 -> "10.0") + Looks for the pattern: current_platform.is_device_capability_family(N) + and converts N (e.g. 100) to a CC string (e.g. "10.x"). """ for node in ast.walk(tree): if not isinstance(node, ast.FunctionDef) or node.name != func_name: @@ -248,15 +247,12 @@ def _find_cc_in_function(tree: ast.AST, func_name: str) -> str | None: if ( isinstance(n, ast.Call) and isinstance(n.func, ast.Attribute) + and n.func.attr == "is_device_capability_family" and n.args and isinstance(n.args[0], ast.Constant) and isinstance(n.args[0].value, int) ): - val = n.args[0].value - if n.func.attr == "is_device_capability_family": - return f"{val // 10}.x" - elif n.func.attr == "is_device_capability": - return f"{val // 10}.{val % 10}" + return f"{n.args[0].value // 10}.x" return None diff --git a/vllm/utils/flashinfer.py b/vllm/utils/flashinfer.py index cd54a06c5a..316816d965 100644 --- a/vllm/utils/flashinfer.py +++ b/vllm/utils/flashinfer.py @@ -305,6 +305,7 @@ def supports_trtllm_attention() -> bool: if envs.VLLM_BATCH_INVARIANT: return False + # Requires SM100 and NVIDIA artifactory to be accessible to download cubins return ( current_platform.is_device_capability_family(100) and has_nvidia_artifactory() )