feat(exl3): online K-quant embedding table (VLLM_EXL3_EMBED_ONLINE_BITS)

docs/features/quantization/online.md

@@ -100,7 +100,7 @@ vllm serve <glm-5.2-modelopt-checkpoint> \
 Serialized checkpoint-quantized shared experts are preserved; only excluded
 BF16 projection weights are converted at load time.
 
-### MXFP8 dense linears on ModelOpt checkpoints
+### MXFP8 dense linears on quantized checkpoints
 
 The `linear` field overlays online MXFP8 the same way onto every other
 BF16 dense linear the ModelOpt checkpoint excludes: attention projections,
@@ -143,6 +143,22 @@ vllm serve lukealonso/GLM-5.2-NVFP4 \
   --quantization-config '{"linear":{"weight":"mxfp8"},"ignore":["re:.*kv_b_proj"]}'
 ```
 
+EXL3 checkpoints can use the same overlay for dense and shared-expert
+projections that are absent from the EXL3 tensor metadata. Serialized EXL3
+dense matrices and routed experts always retain their checkpoint format;
+`moe` overlays are rejected. `lm_head` also remains in its checkpoint format
+or BF16. For example:
+
+```bash
+vllm serve <exl3-checkpoint> \
+  --quantization exl3 \
+  --quantization-config '{"linear":{"weight":"mxfp8"},"shared_experts":{"weight":"mxfp8"},"ignore":["re:.*\\.q_a_proj$","re:.*kv_a_proj_with_mqa","lm_head"]}'
+```
+
+As with ModelOpt, `linear` does not select shared experts. Add the
+`shared_experts` field explicitly when those BF16 projections should also be
+converted. Ignore patterns are matched against unfused checkpoint names.
+
 ### Activation overrides on already-quantized checkpoints
 
 For checkpoint-quantized models, `quantization_config` lets you pick an

tests/models/test_deepseek_v2_mtp_weights.py

@@ -0,0 +1,54 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from types import SimpleNamespace
+
+import pytest
+
+from vllm.model_executor.models.deepseek_v2 import (
+    _skip_disabled_mtp_weight,
+    get_spec_layer_idx_from_weight_name,
+)
+
+
+def test_disabled_mtp_checkpoint_weights_are_skipped():
+    config = SimpleNamespace(num_hidden_layers=78, num_nextn_predict_layers=0)
+
+    assert not _skip_disabled_mtp_weight(config, "model.layers.77.self_attn.weight")
+    assert _skip_disabled_mtp_weight(config, "model.layers.78.self_attn.weight")
+
+
+def test_spec_layer_index_uses_valid_layer_counts():
+    config = SimpleNamespace(num_hidden_layers=78, num_nextn_predict_layers=3)
+
+    assert get_spec_layer_idx_from_weight_name(config, "model.layers.79.weight") == 79
+    assert get_spec_layer_idx_from_weight_name(config, "layers.80.weight") == 80
+    assert get_spec_layer_idx_from_weight_name(config, "model.layers.81.weight") is None
+
+
+def test_spec_layer_index_does_not_scan_configured_layer_count():
+    config = SimpleNamespace(
+        num_hidden_layers=78,
+        num_nextn_predict_layers=10**12,
+    )
+
+    assert get_spec_layer_idx_from_weight_name(config, "model.layers.80.weight") == 80
+    assert (
+        get_spec_layer_idx_from_weight_name(config, "model.embed_tokens.weight") is None
+    )
+
+
+@pytest.mark.parametrize("invalid", [None, True, "3", 3.0, -1, [3]])
+def test_malformed_mtp_layer_counts_fail_closed(invalid):
+    valid = SimpleNamespace(num_hidden_layers=78, num_nextn_predict_layers=3)
+    bad_nextn = SimpleNamespace(num_hidden_layers=78, num_nextn_predict_layers=invalid)
+    bad_hidden = SimpleNamespace(num_hidden_layers=invalid, num_nextn_predict_layers=3)
+
+    for config in (bad_nextn, bad_hidden):
+        assert not _skip_disabled_mtp_weight(config, "model.layers.78.weight")
+        assert (
+            get_spec_layer_idx_from_weight_name(config, "model.layers.78.weight")
+            is None
+        )
+
+    assert get_spec_layer_idx_from_weight_name(valid, "model.layers.78.weight") == 78