Enable quantized attention in NemotronH models (#31898)

Signed-off-by: Roi Koren <roik@nvidia.com>

vllm/model_executor/model_loader/weight_utils.py

@@ -1153,6 +1153,9 @@ def maybe_remap_kv_scale_name(name: str, params_dict: dict) -> str | None:
         # Qwen3 MoE format: .self_attn.qkqkv_proj.{k,v}_scale ->
         # .self_attn.attn.{k,v}_scale
         (r"\.self_attn\.qkqkv_proj\.([kv])_scale$", r".self_attn.attn.\1_scale"),
+        # NemotronH format: .mixer.{k,v}_proj.{k,v}_scale ->
+        # .mixer.attn.{k,v}_scale
+        (r"\.mixer\.[kv]_proj\.([kv])_scale$", r".mixer.attn.\1_scale"),
         # Default format: .{k,v}_scale -> .attn.{k,v}_scale
         (r"\.([kv])_scale$", r".attn.\1_scale"),
     ]

vllm/model_executor/models/nemotron_h.py

@@ -483,6 +483,7 @@ class NemotronHAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             cache_config=cache_config,
+            quant_config=quant_config,
             prefix=f"{prefix}.attn",
         )