[Quantization] Support config.json quantization_config format, fix exclude_modules matching, and fix KV cache scale loading for Nemotron (#18546)

Signed-off-by: root <dafrimi@nvidia.com>
2026-02-21 16:14:29 +08:00
parent e239f8aa85
commit 33c33a7de9
3 changed files with 100 additions and 71 deletions
@@ -61,6 +61,7 @@ from sglang.srt.model_loader.weight_utils import (
    replace_prefix,
    replace_substrings,
 )
+from sglang.srt.models.utils import WeightsMapper
 from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import (
    add_prefix,
@@ -640,6 +641,12 @@ class NemotronHForCausalLM(nn.Module):
        "v_proj.v_scale": "attn.v_scale",
    }

+    hf_to_sglang_mapper = WeightsMapper(
+        orig_to_new_prefix={
+            "backbone.": "model.",
+        }
+    )
+
    def __init__(
        self,
        *,