Route the token embedding table through vLLM's quantized-embedding path
(CompressedTensorsEmbeddingWNA16Int). The kernel exists upstream; the qwen3_5
model code just never passes quant_config to VocabParallelEmbedding. Two
hunks: the main model, and the MTP draft module (without the second one,
single-user mode crashes on load with "no parameter named
'embed_tokens.weight_packed' in Qwen3_5MultiTokenPredictor").

Apply from the installed vllm package directory:
  patch -p1 -d venv/lib/python3.12/site-packages/vllm < patches/qwen3_5-embed-quant.patch

Written against vLLM 0.27.1. Reapply after upgrades.

--- a/model_executor/models/qwen3_5.py
+++ b/model_executor/models/qwen3_5.py
@@ -243,6 +243,8 @@
         self.embed_tokens = VocabParallelEmbedding(
             self.vocab_size,
             config.hidden_size,
+            quant_config=self.quant_config,
+            prefix=maybe_prefix(prefix, "embed_tokens"),
         )

         def get_layer(prefix: str):
--- a/model_executor/models/qwen3_5_mtp.py
+++ b/model_executor/models/qwen3_5_mtp.py
@@ -82,6 +82,8 @@
         self.embed_tokens = VocabParallelEmbedding(
             self.vocab_size,
             config.hidden_size,
+            quant_config=vllm_config.quant_config,
+            prefix=maybe_prefix(prefix, "embed_tokens"),
         )

         # Workaround: mtp.fc is stored as BF16 in NVFP4 checkpoints but is