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, 04:15 UTC
Fix MiniMax-M2 rmsnorm precision and remove useless code (#27627)
Signed-off-by: xuebi <xuebi@minimaxi.com> Co-authored-by: xuebi <xuebi@minimaxi.com>
Roger Young committed
Oct 29, 2025 at 21:01 UTC
d6704dd099b71a9881613dee74b440d78594fa8f
2 files changed
+1
-19
vllm/model_executor/layers/mamba/linear_attn.py
+1
-1
@@ -77,7 +77,7 @@ class MiniMaxText01RMSNormTP(CustomOp):
77
if self.tp_world > 1:
78
variance = tensor_model_parallel_all_reduce(variance) / self.tp_world
79
x = x * torch.rsqrt(variance + self.variance_epsilon)
80
- x = x.to(orig_dtype) * self.weight
80
+ x = (x * self.weight).to(orig_dtype)
81
return x
82
83
def forward(
vllm/model_executor/models/minimax_m2.py
-18
@@ -263,23 +263,6 @@ class MiniMaxM2DecoderLayer(nn.Module):
263
# with the layer's index.
264
layer_idx = int(prefix.split(sep=".")[-1])
265
266
- # TODO: support MTP
267
- attn_window_size = getattr(config, "attn_window_size", None)
268
- if attn_window_size is not None:
269
- if isinstance(attn_window_size, list):
270
- attn_window_size = attn_window_size[layer_idx]
271
- elif isinstance(attn_window_size, int):
272
- attn_window_size = attn_window_size
273
- else:
274
- raise ValueError(f"Invalid attn_window_size: {attn_window_size}")
275
- attn_window_size = None if attn_window_size <= 0 else attn_window_size
276
-
277
- # different rope theta for full layer and swa layer
278
- swa_rope_theta = getattr(config, "swa_rope_theta", -1)
279
- # default to full rope theta
280
- swa_rope_theta = rope_theta if swa_rope_theta <= 0 else swa_rope_theta
281
- rope_theta = swa_rope_theta if attn_window_size is not None else rope_theta
282
-
266
self.layer_idx = layer_idx
267
self.self_attn = MiniMaxM2Attention(
268
hidden_size=self.hidden_size,
@@ -288,7 +271,6 @@ class MiniMaxM2DecoderLayer(nn.Module):
271
rotary_dim=config.rotary_dim,
272
rope_theta=rope_theta,
273
rope_scaling=rope_scaling,
291
- attn_window_size=attn_window_size,
274
max_position_embeddings=max_position_embeddings,
275
rms_norm_eps=config.rms_norm_eps,
276
qkv_bias=getattr(config, "attention_bias", False),