[Bugfix] Restrict TRTLLM attention to SM100, fixing GB300 (SM103) hang (#38730)

Signed-off-by: Stefano Castagnetta <scastagnetta@nvidia.com>
This commit is contained in:
Stefano Castagnetta
2026-04-01 21:08:40 +02:00
committed by GitHub
parent c09ad767cd
commit 6183cae1bd
3 changed files with 14 additions and 10 deletions

View File

@@ -235,10 +235,11 @@ 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_family() calls in a function.
"""Find a compute capability from is_device_capability*() calls in a function.
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").
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")
"""
for node in ast.walk(tree):
if not isinstance(node, ast.FunctionDef) or node.name != func_name:
@@ -247,12 +248,15 @@ 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)
):
return f"{n.args[0].value // 10}.x"
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 None