[model] add qwen3 support (#276)

Author: hiyouga

README.md

@@ -10,7 +10,7 @@ EasyR1 is efficient and scalable due to the design of **[HybirdEngine](https://a
 ## Features
 
 - Supported models
-  - Llama3/Qwen2/Qwen2.5 language models
+  - Llama3/Qwen2/Qwen2.5/Qwen3 language models
   - Qwen2/Qwen2.5-VL vision language models
   - DeepSeek-R1 distill models
 

examples/qwen3_4b_math_grpo.sh

@@ -0,0 +1,12 @@
+#!/bin/bash
+
+set -x
+
+export PYTHONUNBUFFERED=1
+
+MODEL_PATH=Qwen/Qwen3-4B  # replace it with your local file path
+
+python3 -m verl.trainer.main \
+    config=examples/config.yaml \
+    data.max_response_length=4096 \
+    worker.actor.model.model_path=${MODEL_PATH}

verl/models/monkey_patch.py

@@ -20,7 +20,7 @@
 
 
 def apply_ulysses_patch(model_type: str) -> None:
-    if model_type in ("llama", "gemma", "gemma2", "mistral", "qwen2"):
+    if model_type in ("llama", "gemma", "gemma2", "mistral", "qwen2", "qwen3", "qwen3_moe"):
         ALL_ATTENTION_FUNCTIONS["flash_attention_2"] = flash_attention_forward
     elif model_type in ("qwen2_vl", "qwen2_5_vl"):
         from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLFlashAttention2

verl/utils/flops_counter.py

@@ -21,7 +21,7 @@
     from transformers.models.llama.configuration_llama import LlamaConfig
 
 
-VALID_MODLE_TYPE = {"llama", "qwen2", "qwen2_vl", "qwen2_5_vl"}
+VALID_MODLE_TYPE = {"llama", "qwen2", "qwen2_vl", "qwen2_5_vl", "qwen3"}
 
 
 def get_device_flops(unit: str = "T") -> float: