A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Deprecation] Remove DeepGEMM Old Symbol Wrapper (#24902)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Sep 15, 2025 at 22:03 UTC
de2cc3d8678623f0b00be9526db888c1a1496660
1 file changed
+3
-25
vllm/utils/deep_gemm.py
+3
-25
@@ -67,23 +67,6 @@ def _missing(*_: Any, **__: Any) -> NoReturn:
67
"package to enable FP8 kernels.")
68
69
70
-def _resolve_symbol(module, new: str, old: str) -> Callable[..., Any] | None:
71
- """Return the *new* symbol if it exists, otherwise the *old* one."""
72
- if hasattr(module, new):
73
- return getattr(module, new)
74
- if hasattr(module, old):
75
- # TODO(wentao): deprecate old symbol in the future.
76
- logger.warning_once(
77
- "Found legacy DeepGEMM symbol `%s`. Please upgrade the `deep_gemm` "
78
- "package so that `%s` is available. Support for the legacy symbol "
79
- "will be removed in a future vLLM release.",
80
- old,
81
- new,
82
- )
83
- return getattr(module, old)
84
- return None
85
-
86
-
70
_fp8_gemm_nt_impl: Callable[..., Any] | None = None
71
_grouped_impl: Callable[..., Any] | None = None
72
_grouped_masked_impl: Callable[..., Any] | None = None
@@ -109,14 +92,9 @@ def _lazy_init() -> None:
92
93
_dg = importlib.import_module("deep_gemm")
94
112
- _fp8_gemm_nt_impl = _resolve_symbol(_dg, "fp8_gemm_nt",
113
- "gemm_fp8_fp8_bf16_nt")
114
- _grouped_impl = _resolve_symbol(
115
- _dg, "m_grouped_fp8_gemm_nt_contiguous",
116
- "m_grouped_gemm_fp8_fp8_bf16_nt_contiguous")
117
- _grouped_masked_impl = _resolve_symbol(
118
- _dg, "fp8_m_grouped_gemm_nt_masked",
119
- "m_grouped_gemm_fp8_fp8_bf16_nt_masked")
95
+ _fp8_gemm_nt_impl = getattr(_dg, "fp8_gemm_nt", None)
96
+ _grouped_impl = getattr(_dg, "m_grouped_fp8_gemm_nt_contiguous", None)
97
+ _grouped_masked_impl = getattr(_dg, "fp8_m_grouped_gemm_nt_masked", None)
98
99
100
def fp8_gemm_nt(*args, **kwargs):