Fix NVFP4 quantization metadata: stale per-group mxfp4 format, glob-style exclude_modules

The generated config still described the MXFP4 source in three places, each of which makes the checkpoint unloadable:

1. `config_groups.group_0.format` was still `mxfp4-pack-quantized`. SGLang's Kimi-K3 argument override keys on this and forces `moe_runner_backend=flashinfer_mxfp4`, i.e. MXFP4 kernels over NVFP4 tensors.
2. `ignore` used compressed-tensors `re:` regexes. ModelOpt matches with `re.fullmatch(pattern.replace('.', r'\.').replace('*', '.*'))`, so those patterns never match; and it has no `exclude_modules` key at all. Since ModelOpt quantizes everything *except* `exclude_modules` (the inverse of compressed-tensors), unquantized modules were loaded as NVFP4 and failed a shape assertion.
3. `weights` still carried MXFP4 block/scale metadata.

`exclude_modules` is now derived from the modules that actually kept an unpacked 2-D float `.weight` in the shards (2,398 of them, collapsed to 55 globs), verified to cover all of them while excluding none of the 247,296 quantized modules.

config.json

@@ -202,7 +202,7 @@
     "quantization_config": {
       "config_groups": {
         "group_0": {
-          "format": "mxfp4-pack-quantized",
+          "format": "nvfp4-pack-quantized",
           "input_activations": null,
           "output_activations": null,
           "targets": [
@@ -227,19 +227,125 @@
       "format": "nvfp4-pack-quantized",
       "global_compression_ratio": null,
       "ignore": [
-        "re:.*self_attn.*",
-        "re:.*shared_experts.*",
-        "re:.*mlp\\.(gate|up|gate_up|down)_proj.*",
-        "re:.*lm_head.*",
-        "re:.*vision_tower.*",
-        "re:.*mm_projector.*"
+        "*self_attn*",
+        "*shared_experts*",
+        "*mlp.gate_proj*",
+        "*mlp.up_proj*",
+        "*mlp.gate_up_proj*",
+        "*mlp.down_proj*",
+        "*lm_head*",
+        "*vision_tower*",
+        "*mm_projector*",
+        "*language_model.lm_head*",
+        "*language_model.model.embed_tokens*",
+        "*input_layernorm*",
+        "*mlp_res_norm*",
+        "*mlp_res_proj*",
+        "*post_attention_layernorm*",
+        "*self_attention_res_norm*",
+        "*self_attention_res_proj*",
+        "*self_attn.b_proj*",
+        "*self_attn.f_a_proj*",
+        "*self_attn.f_b_proj*",
+        "*self_attn.g_proj*",
+        "*self_attn.k_conv1d*",
+        "*self_attn.k_proj*",
+        "*self_attn.o_norm*",
+        "*self_attn.o_proj*",
+        "*self_attn.q_conv1d*",
+        "*self_attn.q_proj*",
+        "*self_attn.v_conv1d*",
+        "*self_attn.v_proj*",
+        "*block_sparse_moe.gate*",
+        "*block_sparse_moe.routed_expert_down_proj*",
+        "*block_sparse_moe.routed_expert_norm*",
+        "*block_sparse_moe.routed_expert_up_proj*",
+        "*block_sparse_moe.shared_experts.down_proj*",
+        "*block_sparse_moe.shared_experts.gate_proj*",
+        "*block_sparse_moe.shared_experts.up_proj*",
+        "*self_attn.kv_a_layernorm*",
+        "*self_attn.kv_a_proj_with_mqa*",
+        "*self_attn.kv_b_proj*",
+        "*self_attn.q_a_layernorm*",
+        "*self_attn.q_a_proj*",
+        "*self_attn.q_b_proj*",
+        "*language_model.model.norm*",
+        "*language_model.model.output_attn_res_norm*",
+        "*language_model.model.output_attn_res_proj*",
+        "*mm_projector.post_norm*",
+        "*mlp.fc0*",
+        "*mlp.fc1*",
+        "*norm0*",
+        "*norm1*",
+        "*wo*",
+        "*wqkv*",
+        "*vision_tower.encoder.final_layernorm*",
+        "*vision_tower.patch_embed.pos_emb*",
+        "*vision_tower.patch_embed.proj*"
       ],
       "kv_cache_scheme": null,
       "quant_method": "modelopt",
       "quantization_status": "compressed",
       "group_size": 16,
       "quant_algo": "NVFP4",
-      "kv_cache_quant_algo": null
+      "kv_cache_quant_algo": null,
+      "exclude_modules": [
+        "*self_attn*",
+        "*shared_experts*",
+        "*mlp.gate_proj*",
+        "*mlp.up_proj*",
+        "*mlp.gate_up_proj*",
+        "*mlp.down_proj*",
+        "*lm_head*",
+        "*vision_tower*",
+        "*mm_projector*",
+        "*language_model.lm_head*",
+        "*language_model.model.embed_tokens*",
+        "*input_layernorm*",
+        "*mlp_res_norm*",
+        "*mlp_res_proj*",
+        "*post_attention_layernorm*",
+        "*self_attention_res_norm*",
+        "*self_attention_res_proj*",
+        "*self_attn.b_proj*",
+        "*self_attn.f_a_proj*",
+        "*self_attn.f_b_proj*",
+        "*self_attn.g_proj*",
+        "*self_attn.k_conv1d*",
+        "*self_attn.k_proj*",
+        "*self_attn.o_norm*",
+        "*self_attn.o_proj*",
+        "*self_attn.q_conv1d*",
+        "*self_attn.q_proj*",
+        "*self_attn.v_conv1d*",
+        "*self_attn.v_proj*",
+        "*block_sparse_moe.gate*",
+        "*block_sparse_moe.routed_expert_down_proj*",
+        "*block_sparse_moe.routed_expert_norm*",
+        "*block_sparse_moe.routed_expert_up_proj*",
+        "*block_sparse_moe.shared_experts.down_proj*",
+        "*block_sparse_moe.shared_experts.gate_proj*",
+        "*block_sparse_moe.shared_experts.up_proj*",
+        "*self_attn.kv_a_layernorm*",
+        "*self_attn.kv_a_proj_with_mqa*",
+        "*self_attn.kv_b_proj*",
+        "*self_attn.q_a_layernorm*",
+        "*self_attn.q_a_proj*",
+        "*self_attn.q_b_proj*",
+        "*language_model.model.norm*",
+        "*language_model.model.output_attn_res_norm*",
+        "*language_model.model.output_attn_res_proj*",
+        "*mm_projector.post_norm*",
+        "*mlp.fc0*",
+        "*mlp.fc1*",
+        "*norm0*",
+        "*norm1*",
+        "*wo*",
+        "*wqkv*",
+        "*vision_tower.encoder.final_layernorm*",
+        "*vision_tower.patch_embed.pos_emb*",
+        "*vision_tower.patch_embed.proj*"
+      ]
     },
     "remove_invalid_values": false,
     "repetition_penalty": 1.0,

hf_quant_config.json

@@ -8,13 +8,63 @@
     "kv_cache_quant_algo": null,
     "group_size": 16,
     "exclude_modules": [
-      "re:.*lm_head.*",
-      "re:.*mlp\\.(gate|up|gate_up|down)_proj.*",
-      "re:.*mm_projector.*",
-      "re:.*self_attn.*",
-      "re:.*shared_experts.*",
-      "re:.*vision_tower.*"
+      "*self_attn*",
+      "*shared_experts*",
+      "*mlp.gate_proj*",
+      "*mlp.up_proj*",
+      "*mlp.gate_up_proj*",
+      "*mlp.down_proj*",
+      "*lm_head*",
+      "*vision_tower*",
+      "*mm_projector*",
+      "*language_model.lm_head*",
+      "*language_model.model.embed_tokens*",
+      "*input_layernorm*",
+      "*mlp_res_norm*",
+      "*mlp_res_proj*",
+      "*post_attention_layernorm*",
+      "*self_attention_res_norm*",
+      "*self_attention_res_proj*",
+      "*self_attn.b_proj*",
+      "*self_attn.f_a_proj*",
+      "*self_attn.f_b_proj*",
+      "*self_attn.g_proj*",
+      "*self_attn.k_conv1d*",
+      "*self_attn.k_proj*",
+      "*self_attn.o_norm*",
+      "*self_attn.o_proj*",
+      "*self_attn.q_conv1d*",
+      "*self_attn.q_proj*",
+      "*self_attn.v_conv1d*",
+      "*self_attn.v_proj*",
+      "*block_sparse_moe.gate*",
+      "*block_sparse_moe.routed_expert_down_proj*",
+      "*block_sparse_moe.routed_expert_norm*",
+      "*block_sparse_moe.routed_expert_up_proj*",
+      "*block_sparse_moe.shared_experts.down_proj*",
+      "*block_sparse_moe.shared_experts.gate_proj*",
+      "*block_sparse_moe.shared_experts.up_proj*",
+      "*self_attn.kv_a_layernorm*",
+      "*self_attn.kv_a_proj_with_mqa*",
+      "*self_attn.kv_b_proj*",
+      "*self_attn.q_a_layernorm*",
+      "*self_attn.q_a_proj*",
+      "*self_attn.q_b_proj*",
+      "*language_model.model.norm*",
+      "*language_model.model.output_attn_res_norm*",
+      "*language_model.model.output_attn_res_proj*",
+      "*mm_projector.post_norm*",
+      "*mlp.fc0*",
+      "*mlp.fc1*",
+      "*norm0*",
+      "*norm1*",
+      "*wo*",
+      "*wqkv*",
+      "*vision_tower.encoder.final_layernorm*",
+      "*vision_tower.patch_embed.pos_emb*",
+      "*vision_tower.patch_embed.proj*"
     ],
-    "quantized_modules_count": 247296
+    "quantized_modules_count": 247296,
+    "activation_scheme": "none"
   }
 }