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 10, 12:45 UTC
Add back missing DeepEP LL params (#31911)
Signed-off-by: Elvir Crncevic <elvircrn@gmail.com>
Elvir Crnčević committed
Jan 7, 2026 at 23:47 UTC
ffc0a2798b118f7ceb21645df59d2bfdfc461d42
1 file changed
+2
vllm/model_executor/layers/fused_moe/deepep_ll_prepare_finalize.py
+2
@@ -295,6 +295,8 @@ class DeepEPLLPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
295
self.max_tokens_per_rank,
296
num_experts,
297
use_fp8=self.use_fp8_dispatch,
298
+ round_scale=self.use_ue8m0_dispatch,
299
+ use_ue8m0=self.use_ue8m0_dispatch,
300
**(dict(use_nvfp4=True) if use_nvfp4 else dict()),
301
**(
302
dict(x_global_scale=qc_a1_gscale_or_scale)