[Model] Add MiniMax M3 support (#45381)

Signed-off-by: youkaichao <youkaichao@gmail.com>
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Signed-off-by: functionstackx <47992694+functionstackx@users.noreply.github.com>
Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Co-authored-by: Thien Tran <gau.nernst@yahoo.com.sg>
Co-authored-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>
Co-authored-by: Roger Wang <hey@rogerw.io>
Co-authored-by: functionstackx <47992694+functionstackx@users.noreply.github.com>
Co-authored-by: Yongye Zhu <zyy1102000@gmail.com>
Co-authored-by: Jee Jee Li <jeejeelee@inferact.ai>
This commit is contained in:
youkaichao
2026-06-16 01:01:25 +08:00
committed by GitHub
co-authored by OpenAI Codex Isotr0py Thien Tran Bugen Zhao Jee Jee Li Roger Wang functionstackx Yongye Zhu Jee Jee Li
parent a3195fab7b
commit 0a1c5034f5
108 changed files with 14734 additions and 311 deletions
+10
View File
@@ -113,6 +113,7 @@ if TYPE_CHECKING:
VLLM_ENABLE_FLA_PACKED_RECURRENT_DECODE: bool = True
VLLM_DISABLE_PYNCCL: bool = False
VLLM_USE_OINK_OPS: bool = False
VLLM_MXFP8_EMULATION_DEQUANT_AT_LOAD: bool = True
VLLM_ROCM_USE_AITER: bool = False
VLLM_ROCM_USE_AITER_PAGED_ATTN: bool = False
VLLM_ROCM_USE_AITER_LINEAR: bool = True
@@ -1092,6 +1093,15 @@ environment_variables: dict[str, Callable[[], Any]] = {
),
# Disable aiter ops unless specifically enabled.
# Acts as a parent switch to enable the rest of the other operations.
# On hardware without a native MXFP8 kernel (e.g. ROCm gfx942 / MI300), the
# MXFP8 emulation path dequantizes weights MXFP8->BF16 once at load time and
# runs as a BF16 checkpoint (no per-step dequant). Set to 0 to fall back to
# per-step dequant: keeps the 1-byte MXFP8 weights (~half the weight memory)
# at the cost of dequantizing every forward step (much slower). Default on.
"VLLM_MXFP8_EMULATION_DEQUANT_AT_LOAD": lambda: (
os.getenv("VLLM_MXFP8_EMULATION_DEQUANT_AT_LOAD", "True").lower()
in ("true", "1")
),
"VLLM_ROCM_USE_AITER": lambda: (
os.getenv("VLLM_ROCM_USE_AITER", "False").lower() in ("true", "1")
),