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 8, 21:30 UTC
[Bugfix] Kimi-K2 grouped_topk usage for Flashinfer monolithic kernels. (#33858)
Signed-off-by: Pavani Majety <pmajety@nvidia.com>
Pavani Majety committed
Feb 5, 2026 at 01:32 UTC
d2f4a71cd54418369f617a174e6c839a71a47ed8
1 file changed
+3
-11
vllm/model_executor/models/deepseek_v2.py
+3
-11
@@ -295,14 +295,6 @@ class DeepseekV2MoE(nn.Module):
295
prefix=f"{prefix}.shared_experts",
296
)
297
298
- n_group = getattr(config, "n_group", 1)
299
- topk_group = getattr(config, "topk_group", 1)
300
- use_grouped_topk = True
301
- if (n_group, topk_group) == (1, 1):
302
- n_group = None
303
- topk_group = None
304
- use_grouped_topk = False
305
-
298
self.experts = SharedFusedMoE(
299
shared_experts=self.shared_experts,
300
gate=self.gate,
@@ -313,9 +305,9 @@ class DeepseekV2MoE(nn.Module):
305
reduce_results=False,
306
renormalize=config.norm_topk_prob,
307
quant_config=quant_config,
316
- use_grouped_topk=use_grouped_topk,
317
- num_expert_group=n_group,
318
- topk_group=topk_group,
308
+ use_grouped_topk=True,
309
+ num_expert_group=getattr(config, "n_group", 1),
310
+ topk_group=getattr(config, "topk_group", 1),
311
prefix=f"{prefix}.experts",
312
scoring_func=getattr(config, "scoring_func", "softmax"),
313
# we do scaling outside, set factor to 1.0 to avoid double mul