fix qwen 2.5 vl

Signed-off-by: Dylan Chen <[email protected]>

Author: DylanChen-NV

examples/llm_ptq/example_utils.py

@@ -56,7 +56,7 @@ def get_tokenizer(ckpt_path, trust_remote_code=False, **kwargs):
         ckpt_path, trust_remote_code=trust_remote_code, **kwargs
     )
 
-    if "qwen" in type(tokenizer).__name__.lower():
+    if "qwen" in type(tokenizer).__name__.lower() and "vl" not in ckpt_path.lower():
         # qwen use token id 151643 as pad and eos tokens
         tokenizer.pad_token = tokenizer.convert_ids_to_tokens(151643)
         tokenizer.eos_token = tokenizer.convert_ids_to_tokens(151643)

examples/vlm_ptq/scripts/huggingface_example.sh

@@ -145,6 +145,10 @@ case "${MODEL_TYPE}" in
         VISUAL_FEATURE=576
         VLM_ARGS=" --max_multimodal_len=$((BUILD_MAX_BATCH_SIZE * VISUAL_FEATURE)) "
         ;;
+    "qwen")
+        VISUAL_FEATURE=1280
+        VLM_ARGS=" --max_multimodal_len=$((BUILD_MAX_BATCH_SIZE * VISUAL_FEATURE)) "
+        ;;
     "mllama")
         PTQ_ARGS+=" --kv_cache_qformat none "
         VLM_ARGS=" --max_encoder_input_len=6404 --skip_run"