From 03cf2de2e3601f9ccbdf64e02e51dc1f77fe396e Mon Sep 17 00:00:00 2001 From: YAMY <74099316+YAMY1234@users.noreply.github.com> Date: Wed, 19 Aug 2026 11:12:32 -0700 Subject: [PATCH] [Qwen3.5][MTP] Preserve online NVFP4 draft quantization for mixed checkpoints (#35545) --- python/sglang/srt/configs/model_config.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/python/sglang/srt/configs/model_config.py b/python/sglang/srt/configs/model_config.py index 0b55e81ba..1afb405c2 100644 --- a/python/sglang/srt/configs/model_config.py +++ b/python/sglang/srt/configs/model_config.py @@ -1516,14 +1516,14 @@ class ModelConfig: "quant_method", "" if not self.quantization else self.quantization ).lower() - # ModelOpt FP4 checkpoints quantize only the target model; an - # embedded MTP draft may stay unquantized, so an explicit + # ModelOpt FP4 and mixed checkpoints can quantize only the target + # model; an embedded MTP draft may stay unquantized, so an explicit # nvfp4_online opt-in for the draft wins over checkpoint detection. # The online loader rejects already-packed weights at load time. preserve_online_draft_quantization = ( self.is_draft_model and self.quantization == "nvfp4_online" - and quant_method == "modelopt_fp4" + and quant_method in ("modelopt_fp4", "modelopt_mixed") ) # An explicit online-requantization request (e.g. quark_mxfp4 on top # of an NVFP4/mixed checkpoint) must not be overridden back to the