qwen3 asr working

Author: ngxson

convert_hf_to_gguf.py

@@ -5057,6 +5057,14 @@ def set_vocab(self):
         super().set_vocab()
         # fix chat template, use correct chatml format
         self.gguf_writer.add_chat_template("{% for message in messages %}{{'<|im_start|>' + message['role'] + '\\n' + message['content'] + '<|im_end|>' + '\\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\\n' }}{% endif %}")
+        # correct BOS/EOS tokens
+        with open(self.dir_model / "tokenizer_config.json", "r", encoding="utf-8") as f:
+            tokenizer_config = json.load(f)
+            added_tokens = tokenizer_config.get("added_tokens_decoder", {})
+            for token_id, data in added_tokens.items():
+                if data.get("content") == "<|im_end|>":
+                    self.gguf_writer.add_bos_token_id(int(token_id))
+                    self.gguf_writer.add_eos_token_id(int(token_id))
 
     def modify_tensors(self, data_torch, name, bid):
         # qwen3-omni

tools/mtmd/mtmd.cpp

@@ -982,6 +982,10 @@ bool mtmd_decode_use_non_causal(mtmd_context * ctx) {
 }
 
 bool mtmd_decode_use_mrope(mtmd_context * ctx) {
+    if (ctx->ctx_v == nullptr && ctx->proj_type_a() == PROJECTOR_TYPE_QWEN3A) {
+        // qwen3-asr
+        return true;
+    }
     switch (ctx->proj_type_v()) {
         case PROJECTOR_TYPE_QWEN2VL:
         case PROJECTOR_TYPE_QWEN25VL: