Bumping version to 0.5.2

README.md

@@ -32,8 +32,8 @@ Trinity-RFT provides functionalities for users with different backgrounds and ob
 
 ## 🚀 News
 
+* [2026-04] [[Release Notes]](https://github.com/agentscope-ai/Trinity-RFT/releases/tag/v0.5.2) Trinity-RFT v0.5.2 released: support for Qwen3.5 series, bug fixes and optimizations.
 * [2026-03] 🤖 Trinity-RFT empowers the training of CoPaw-Flash, building a small agent model better suited for localized scenarios. Feel free to try CoPaw-Flash on [CoPaw](https://github.com/agentscope-ai/CoPaw); models are also available on [ModelScope](https://www.modelscope.cn/organization/AgentScope) and [HuggingFace](https://huggingface.co/agentscope-ai) ([News](https://mp.weixin.qq.com/s/-BXNU_PMi6QJuwSB5BqTbQ)).
-* [2026-03] Trinity-RFT now supports Qwen3.5 series.
 * [2026-02] [[Release Notes]](https://github.com/agentscope-ai/Trinity-RFT/releases/tag/v0.5.1) Trinity-RFT v0.5.1 released: Enhanced VLM support, logging improvements, bug fixes.
 * [2026-02] [[Release Notes]](https://github.com/agentscope-ai/Trinity-RFT/releases/tag/v0.5.0) Trinity-RFT v0.5.0 released: colocate mode for single-GPU scenarios, trainer driven weight synchronization, automatic parallelism setting suggestion, and more.
 * [2026-01] 🎉 Three papers accepted by ICLR 2026: [CHORD](https://github.com/agentscope-ai/Trinity-RFT/tree/main/examples/mix_chord), [BOTS](https://github.com/agentscope-ai/Trinity-RFT/tree/main/examples/bots), and [Group-relative REINFORCE variants](https://github.com/agentscope-ai/Trinity-RFT/tree/main/examples/rec_gsm8k). Try out these new algorithms in Trinity-RFT!
@@ -47,7 +47,7 @@ Trinity-RFT provides functionalities for users with different backgrounds and ob
 <ul>
   <li> [2025-11] Introducing [BOTS](https://github.com/agentscope-ai/Trinity-RFT/tree/main/examples/bots): online RL task selection for efficient LLM fine-tuning ([paper](https://arxiv.org/pdf/2510.26374)).</li>
   <li> [2025-09] [Our paper](https://arxiv.org/pdf/2509.24203) reveals a novel off-policy interpretation for group-relative REINFORCE and its variants like GRPO and AsymRE ([implementation](https://github.com/agentscope-ai/Trinity-RFT/tree/main/examples/rec_gsm8k)).</li>
-  <li> [2026-03] [2025-08] Introducing [CHORD](https://github.com/agentscope-ai/Trinity-RFT/tree/main/examples/mix_chord): dynamic SFT + RL integration for advanced LLM fine-tuning ([paper](https://arxiv.org/pdf/2508.11408)).</li>
+  <li> [2025-08] Introducing [CHORD](https://github.com/agentscope-ai/Trinity-RFT/tree/main/examples/mix_chord): dynamic SFT + RL integration for advanced LLM fine-tuning ([paper](https://arxiv.org/pdf/2508.11408)).</li>
   <li> [2025-11] Trinity-RFT v0.3.3 released: bug fixes.</li>
   <li> [2025-11] Trinity-RFT v0.3.2 released: bug fixes and advanced task selection & scheduling.</li>
   <li> [2025-10] Trinity-RFT v0.3.1 released: multi-stage training support, improved agentic RL examples, LoRA support, debug mode and new RL algorithms.</li>

README_zh.md

@@ -41,8 +41,8 @@ Trinity-RFT 面向不同背景和目标的用户提供相应功能：
 
 ## 🚀 新闻
 
+* [2026-04] [[发布说明]](https://github.com/agentscope-ai/Trinity-RFT/releases/tag/v0.5.2) Trinity-RFT v0.5.2 发布：支持 Qwen3.5 系列，修复 Bug 并进行多项优化。
 * [2026-03] 🤖 Trinity-RFT 助力 CoPaw-Flash 训练，打造更懂本地化场景的智能体小模型。欢迎到 [CoPaw](https://github.com/agentscope-ai/CoPaw) 试用 CoPaw-Flash，模型下载请见 [ModelScope](https://www.modelscope.cn/organization/AgentScope) 和 [HuggingFace](https://huggingface.co/agentscope-ai)（[新闻](https://mp.weixin.qq.com/s/-BXNU_PMi6QJuwSB5BqTbQ)）。
-* [2026-03] Trinity-RFT 现在已经支持 Qwen3.5 系列模型。
 * [2026-02] [[发布说明]](https://github.com/agentscope-ai/Trinity-RFT/releases/tag/v0.5.1) Trinity-RFT v0.5.1 发布：增强 VLM 支持，改进日志系统，修复若干 Bug。
 * [2026-02] [[发布说明]](https://github.com/agentscope-ai/Trinity-RFT/releases/tag/v0.5.0) Trinity-RFT v0.5.0 发布：单 GPU 场景下的 colocate 模式，trainer 驱动的权重同步，自动并行设置建议等新功能。
 * [2026-01] 🎉 三篇论文被 ICLR 2026 接收：[CHORD](https://github.com/agentscope-ai/Trinity-RFT/tree/main/examples/mix_chord)、[BOTS](https://github.com/agentscope-ai/Trinity-RFT/tree/main/examples/bots) 和 [Group-relative REINFORCE 系列变种](https://github.com/agentscope-ai/Trinity-RFT/tree/main/examples/rec_gsm8k)。在 Trinity-RFT 中尝试这些新算法吧！

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name =  "trinity-rft"
-version = "0.5.1"
+dynamic = ["version"]
 authors = [
     {name="Trinity-RFT Team", email="trinity-rft@outlook.com"},
 ]
@@ -125,6 +125,9 @@ exclude = ["tests*", "docs*", "scripts*"]
 [tool.setuptools.package-data]
 my_package = ["*.md", "*.rst"]
 
+[tool.setuptools.dynamic]
+version = {attr = "trinity.__version__"}
+
 [tool.black]
 line-length = 100
 target-version = ["py310", "py311", "py312"]

tests/common/vllm_test.py

@@ -203,13 +203,24 @@ def _check_experience(exp):
             self.assertEqual(len(encoded_prompt), exp.prompt_length)
             self.assertLessEqual(exp.prompt_length, self.config.model.max_prompt_tokens)
             # check response content and length
-            encoded_response = self.tokenizer.encode(exp.response_text, add_special_tokens=False)
-            self.assertEqual(len(encoded_response), len(exp.tokens) - exp.prompt_length)
-            self.assertLessEqual(
-                len(exp.tokens) - exp.prompt_length, self.config.model.max_response_tokens
-            )
-            # check full sequence
-            self.assertLessEqual(len(exp.tokens), self.config.model.max_model_len)
+            if exp.truncate_status == "prompt_truncated":
+                self.assertEqual(
+                    exp.response_text, "[This experience is masked out due to overlong prompt]"
+                )
+                self.assertEqual(exp.prompt_text, self.tokenizer.decode(exp.tokens[:-1]))
+                self.assertEqual(len(exp.tokens), self.config.model.max_prompt_tokens + 1)
+                self.assertEqual(exp.prompt_length, self.config.model.max_prompt_tokens)
+                self.assertTrue(torch.equal(exp.logprobs, torch.zeros(1, dtype=torch.float32)))
+            else:
+                encoded_response = self.tokenizer.encode(
+                    exp.response_text, add_special_tokens=False
+                )
+                self.assertEqual(len(encoded_response), len(exp.tokens) - exp.prompt_length)
+                self.assertLessEqual(
+                    len(exp.tokens) - exp.prompt_length, self.config.model.max_response_tokens
+                )
+                # check full sequence
+                self.assertLessEqual(len(exp.tokens), self.config.model.max_model_len)
 
         # For vllm engine, max_prompt_tokens and max_response_tokens work
         response = self.model_wrapper.chat(messages)
@@ -241,6 +252,30 @@ def _check_experience(exp):
             self.config.model.max_response_tokens,
         )
 
+        # test prompt truncation branch in generate
+        if self.max_prompt_tokens == 5:
+            await prepare_engines(self.engines, self.auxiliary_engines)
+            await self.model_wrapper.prepare()
+
+            prompt = "This is a deliberately long prompt for truncation coverage."
+            prompt_token_ids = self.tokenizer(prompt, truncation=False, return_tensors="pt")[
+                "input_ids"
+            ][0].tolist()
+            self.assertGreater(len(prompt_token_ids), self.config.model.max_prompt_tokens)
+
+            responses = self.model_wrapper.generate([prompt], n=2)
+            self.assertEqual(len(responses), 2)
+
+            for response in responses:
+                self.assertEqual(response.truncate_status, "prompt_truncated")
+                _check_experience(response)
+
+            exps = self.model_wrapper.extract_experience_from_history()
+            self.assertEqual(len(exps), 2)
+            for exp in exps:
+                self.assertEqual(exp.truncate_status, "prompt_truncated")
+                _check_experience(exp)
+
 
 class TestModelLenWithoutPromptTruncation(RayUnittestBaseAsync):
     def setUp(self):

trinity/__init__.py

@@ -1,4 +1,4 @@
 # -*- coding: utf-8 -*-
 """Trinity-RFT (Reinforcement Fine-Tuning)"""
 
-__version__ = "0.5.1"
+__version__ = "0.5.2"

trinity/common/models/model.py

@@ -142,20 +142,14 @@ def _handle_prompt_truncation(self, prompt: str, **kwargs) -> Tuple[Sequence, bo
             self.logger.warning(f"Prompt was truncated to {self.config.max_prompt_tokens} tokens")
 
             dummy_response = "[This experience is masked out due to overlong prompt]"
-            dummy_response_tokens = self.tokenizer(  # type: ignore
-                dummy_response, truncation=False, return_tensors="pt"
-            )["input_ids"][0].tolist()
-            dummy_response_tokens = dummy_response_tokens[
-                : min(len(dummy_response_tokens), self.config.max_response_tokens)  # type: ignore
-            ]
-
-            token_ids = prompt_token_ids[: self.config.max_prompt_tokens] + dummy_response_tokens
+
+            token_ids = prompt_token_ids[: self.config.max_prompt_tokens + 1]
             return [
                 Experience(
                     tokens=token_ids,
-                    logprobs=torch.zeros(len(dummy_response_tokens), dtype=torch.float32),
-                    prompt_length=len(prompt_token_ids),
-                    prompt_text=self.tokenizer.decode(prompt_token_ids),
+                    logprobs=torch.zeros(1, dtype=torch.float32),
+                    prompt_length=self.config.max_prompt_tokens,  # Use truncated length
+                    prompt_text=self.tokenizer.decode(token_ids[:-1]),
                     response_text=dummy_response,
                     truncate_status="prompt_truncated",
                     reward=0.0,