Multi-turn tool calling support

docs/source/experimental.md

@@ -66,7 +66,7 @@ class GroupFilter:
         return group_scores
 
 training_args = GFPOConfig(
-    output_dir="Qwen3-0.6B-GFPO"
+    output_dir="Qwen3-0.6B-GFPO",
     per_device_train_batch_size=4,
     num_remains_in_group=2,
     bf16=True,

scripts/generate_tiny_models.py

@@ -292,6 +292,7 @@ def init_weights_tiny_model(model):
         "hidden_size": 16,
         "num_attention_heads": 4,
         "num_key_value_heads": 2,
+        "embed_dim": 64,
     }
     config = AutoConfig.from_pretrained(model_id, text_config=text_config, vision_config=vision_config)
 

tests/test_grpo_trainer.py

@@ -1258,6 +1258,10 @@ def test_prepare_input_called_with_correct_data(self):
     def test_training_vlm(self, model_id):
         dataset = load_dataset("trl-internal-testing/zen-image", "conversational_prompt_only", split="train")
 
+        def reward_func(completions, **kwargs):
+            """Reward function that rewards longer completions."""
+            return [float(len(completion[0]["content"])) for completion in completions]
+
         training_args = GRPOConfig(
             output_dir=self.tmp_dir,
             learning_rate=0.1,  # increase the learning rate to speed up the test
@@ -1269,7 +1273,7 @@ def test_training_vlm(self, model_id):
         )
         trainer = GRPOTrainer(
             model=model_id,
-            reward_funcs="trl-internal-testing/tiny-Qwen2ForSequenceClassification-2.5",
+            reward_funcs=reward_func,
             args=training_args,
             train_dataset=dataset,
         )
@@ -1301,6 +1305,10 @@ def test_training_vlm(self, model_id):
     def test_training_vlm_beta_non_zero(self):
         dataset = load_dataset("trl-internal-testing/zen-image", "conversational_prompt_only", split="train")
 
+        def reward_func(completions, **kwargs):
+            """Reward function that rewards longer completions."""
+            return [float(len(completion[0]["content"])) for completion in completions]
+
         training_args = GRPOConfig(
             output_dir=self.tmp_dir,
             beta=0.1,  # set beta to non-zero value to test the case where the reference model is used
@@ -1312,7 +1320,7 @@ def test_training_vlm_beta_non_zero(self):
         )
         trainer = GRPOTrainer(
             model="trl-internal-testing/tiny-Qwen2_5_VLForConditionalGeneration",
-            reward_funcs="trl-internal-testing/tiny-Qwen2ForSequenceClassification-2.5",
+            reward_funcs=reward_func,
             args=training_args,
             train_dataset=dataset,
         )
@@ -1340,7 +1348,11 @@ def test_training_vlm_peft(self):
             "trl-internal-testing/tiny-Qwen2_5_VLForConditionalGeneration"
         )
         base_param_names = [f"base_model.model.{n}" for n, _ in model.named_parameters()]
-        dataset = load_dataset("trl-internal-testing/zen", "standard_prompt_only", split="train")
+        dataset = load_dataset("trl-internal-testing/zen-image", "conversational_prompt_only", split="train")
+
+        def reward_func(completions, **kwargs):
+            """Reward function that rewards longer completions."""
+            return [float(len(completion[0]["content"])) for completion in completions]
 
         training_args = GRPOConfig(
             output_dir=self.tmp_dir,
@@ -1352,7 +1364,7 @@ def test_training_vlm_peft(self):
         )
         trainer = GRPOTrainer(
             model=model,
-            reward_funcs="trl-internal-testing/tiny-Qwen2ForSequenceClassification-2.5",
+            reward_funcs=reward_func,
             args=training_args,
             train_dataset=dataset,
             peft_config=LoraConfig(target_modules=["q_proj", "v_proj"]),
@@ -1376,6 +1388,10 @@ def test_training_vlm_peft(self):
     def test_training_vlm_and_importance_sampling(self):
         dataset = load_dataset("trl-internal-testing/zen-image", "conversational_prompt_only", split="train")
 
+        def reward_func(completions, **kwargs):
+            """Reward function that rewards longer completions."""
+            return [float(len(completion[0]["content"])) for completion in completions]
+
         training_args = GRPOConfig(
             output_dir=self.tmp_dir,
             learning_rate=0.1,  # increase the learning rate to speed up the test
@@ -1387,7 +1403,7 @@ def test_training_vlm_and_importance_sampling(self):
         )
         trainer = GRPOTrainer(
             model="trl-internal-testing/tiny-Qwen2_5_VLForConditionalGeneration",
-            reward_funcs="trl-internal-testing/tiny-Qwen2ForSequenceClassification-2.5",
+            reward_funcs=reward_func,
             args=training_args,
             train_dataset=dataset,
         )
@@ -1413,6 +1429,10 @@ def test_training_vlm_and_importance_sampling(self):
     def test_training_vlm_and_liger(self):
         dataset = load_dataset("trl-internal-testing/zen-image", "conversational_prompt_only", split="train")
 
+        def reward_func(completions, **kwargs):
+            """Reward function that rewards longer completions."""
+            return [float(len(completion[0]["content"])) for completion in completions]
+
         training_args = GRPOConfig(
             output_dir=self.tmp_dir,
             learning_rate=0.1,  # increase the learning rate to speed up the test
@@ -1425,7 +1445,7 @@ def test_training_vlm_and_liger(self):
         )
         trainer = GRPOTrainer(
             model="trl-internal-testing/tiny-Qwen2_5_VLForConditionalGeneration",
-            reward_funcs="trl-internal-testing/tiny-Qwen2ForSequenceClassification-2.5",
+            reward_funcs=reward_func,
             args=training_args,
             train_dataset=dataset,
         )
@@ -1451,6 +1471,10 @@ def test_training_vlm_and_prompt_truncation(self):
         # If not handled properly, prompt truncation may truncate image token
         dataset = load_dataset("trl-internal-testing/zen-image", "conversational_prompt_only", split="train")
 
+        def reward_func(completions, **kwargs):
+            """Reward function that rewards longer completions."""
+            return [float(len(completion[0]["content"])) for completion in completions]
+
         training_args = GRPOConfig(
             output_dir=self.tmp_dir,
             learning_rate=0.1,  # increase the learning rate to speed up the test
@@ -1462,7 +1486,7 @@ def test_training_vlm_and_prompt_truncation(self):
         )
         trainer = GRPOTrainer(
             model="trl-internal-testing/tiny-Qwen2_5_VLForConditionalGeneration",
-            reward_funcs="trl-internal-testing/tiny-Qwen2ForSequenceClassification-2.5",
+            reward_funcs=reward_func,
             args=training_args,
             train_dataset=dataset,
         )
@@ -1495,6 +1519,10 @@ def test_training_vlm_and_prompt_truncation(self):
     def test_training_vlm_and_vllm(self, model_id) -> None:
         dataset = load_dataset("trl-internal-testing/zen-image", "conversational_prompt_only", split="train")
 
+        def reward_func(completions, **kwargs):
+            """Reward function that rewards longer completions."""
+            return [float(len(completion[0]["content"])) for completion in completions]
+
         training_args = GRPOConfig(
             output_dir=self.tmp_dir,
             learning_rate=0.1,
@@ -1508,7 +1536,44 @@ def test_training_vlm_and_vllm(self, model_id) -> None:
         )
         trainer = GRPOTrainer(
             model=model_id,
-            reward_funcs="trl-internal-testing/tiny-Qwen2ForSequenceClassification-2.5",
+            reward_funcs=reward_func,
+            args=training_args,
+            train_dataset=dataset,
+        )
+
+        previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
+
+        trainer.train()
+
+        self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
+
+        for n, param in previous_trainable_params.items():
+            new_param = trainer.model.get_parameter(n)
+            self.assertFalse(torch.equal(param, new_param), f"Parameter {n} has not changed.")
+
+    @require_vision
+    def test_training_vlm_multi_image(self):
+        dataset = load_dataset("trl-internal-testing/zen-multi-image", "conversational_prompt_only", split="train")
+
+        # For now, mixing image+text and text-only examples is not supported, so we filter out text-only examples
+        dataset = dataset.filter(lambda x: len(x["images"]) > 0)
+
+        def reward_func(completions, **kwargs):
+            """Reward function that rewards longer completions."""
+            return [float(len(completion[0]["content"])) for completion in completions]
+
+        training_args = GRPOConfig(
+            output_dir=self.tmp_dir,
+            learning_rate=0.1,  # increase the learning rate to speed up the test
+            per_device_train_batch_size=3,  # reduce the batch size to reduce memory usage
+            num_generations=3,  # reduce the number of generations to reduce memory usage
+            max_completion_length=8,  # reduce the completion length to reduce memory usage
+            max_prompt_length=None,  # disable prompt truncation, because usually, models don't support it
+            report_to="none",
+        )
+        trainer = GRPOTrainer(
+            model="trl-internal-testing/tiny-Qwen2_5_VLForConditionalGeneration",
+            reward_funcs=reward_func,
             args=training_args,
             train_dataset=dataset,
         )
@@ -1519,7 +1584,20 @@ def test_training_vlm_and_vllm(self, model_id) -> None:
 
         self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
 
+        # Check that the params have changed
+        # Because of the way the tiny models are initialized, the gradient does not flow properly through the
+        # vision parts of the model, so we skip them. Ideally, we should fix the init of these models.
+        params_to_skip = (
+            # "model.vision_tower.",
+            # "model.multi_modal_projector.",
+            # "model.vision_model.",
+            # "model.connector.modality_projection.",
+            # "model.visual.",
+            # "model.image_newline",
+        )
         for n, param in previous_trainable_params.items():
+            if n.startswith(params_to_skip):
+                continue
             new_param = trainer.model.get_parameter(n)
             self.assertFalse(torch.equal(param, new_param), f"Parameter {n} has not changed.")
 

tests/test_rloo_trainer.py

@@ -1089,6 +1089,10 @@ def test_prepare_input_called_with_correct_data(self):
     def test_training_vlm(self, model_id):
         dataset = load_dataset("trl-internal-testing/zen-image", "conversational_prompt_only", split="train")
 
+        def reward_func(completions, **kwargs):
+            """Reward function that rewards longer completions."""
+            return [float(len(completion[0]["content"])) for completion in completions]
+
         training_args = RLOOConfig(
             output_dir=self.tmp_dir,
             learning_rate=0.1,  # increase the learning rate to speed up the test
@@ -1100,7 +1104,7 @@ def test_training_vlm(self, model_id):
         )
         trainer = RLOOTrainer(
             model=model_id,
-            reward_funcs="trl-internal-testing/tiny-Qwen2ForSequenceClassification-2.5",
+            reward_funcs=reward_func,
             args=training_args,
             train_dataset=dataset,
         )
@@ -1132,6 +1136,10 @@ def test_training_vlm(self, model_id):
     def test_training_vlm_beta_non_zero(self):
         dataset = load_dataset("trl-internal-testing/zen-image", "conversational_prompt_only", split="train")
 
+        def reward_func(completions, **kwargs):
+            """Reward function that rewards longer completions."""
+            return [float(len(completion[0]["content"])) for completion in completions]
+
         training_args = RLOOConfig(
             output_dir=self.tmp_dir,
             beta=0.1,  # set beta to non-zero value to test the case where the reference model is used
@@ -1143,7 +1151,7 @@ def test_training_vlm_beta_non_zero(self):
         )
         trainer = RLOOTrainer(
             model="trl-internal-testing/tiny-Qwen2_5_VLForConditionalGeneration",
-            reward_funcs="trl-internal-testing/tiny-Qwen2ForSequenceClassification-2.5",
+            reward_funcs=reward_func,
             args=training_args,
             train_dataset=dataset,
         )
@@ -1171,7 +1179,11 @@ def test_training_vlm_peft(self):
             "trl-internal-testing/tiny-Qwen2_5_VLForConditionalGeneration"
         )
         base_param_names = [f"base_model.model.{n}" for n, _ in model.named_parameters()]
-        dataset = load_dataset("trl-internal-testing/zen", "standard_prompt_only", split="train")
+        dataset = load_dataset("trl-internal-testing/zen-image", "conversational_prompt_only", split="train")
+
+        def reward_func(completions, **kwargs):
+            """Reward function that rewards longer completions."""
+            return [float(len(completion[0]["content"])) for completion in completions]
 
         training_args = RLOOConfig(
             output_dir=self.tmp_dir,
@@ -1183,7 +1195,7 @@ def test_training_vlm_peft(self):
         )
         trainer = RLOOTrainer(
             model=model,
-            reward_funcs="trl-internal-testing/tiny-Qwen2ForSequenceClassification-2.5",
+            reward_funcs=reward_func,
             args=training_args,
             train_dataset=dataset,
             peft_config=LoraConfig(target_modules=["q_proj", "v_proj"]),
@@ -1208,6 +1220,10 @@ def test_training_vlm_and_prompt_truncation(self):
         # If not handled properly, prompt truncation may truncate image token
         dataset = load_dataset("trl-internal-testing/zen-image", "conversational_prompt_only", split="train")
 
+        def reward_func(completions, **kwargs):
+            """Reward function that rewards longer completions."""
+            return [float(len(completion[0]["content"])) for completion in completions]
+
         training_args = RLOOConfig(
             output_dir=self.tmp_dir,
             learning_rate=0.1,  # increase the learning rate to speed up the test
@@ -1219,7 +1235,7 @@ def test_training_vlm_and_prompt_truncation(self):
         )
         trainer = RLOOTrainer(
             model="trl-internal-testing/tiny-Qwen2_5_VLForConditionalGeneration",
-            reward_funcs="trl-internal-testing/tiny-Qwen2ForSequenceClassification-2.5",
+            reward_funcs=reward_func,
             args=training_args,
             train_dataset=dataset,
         )
@@ -1252,6 +1268,10 @@ def test_training_vlm_and_prompt_truncation(self):
     def test_training_vlm_and_vllm(self, model_id) -> None:
         dataset = load_dataset("trl-internal-testing/zen-image", "conversational_prompt_only", split="train")
 
+        def reward_func(completions, **kwargs):
+            """Reward function that rewards longer completions."""
+            return [float(len(completion[0]["content"])) for completion in completions]
+
         training_args = RLOOConfig(
             output_dir=self.tmp_dir,
             learning_rate=0.1,
@@ -1265,7 +1285,44 @@ def test_training_vlm_and_vllm(self, model_id) -> None:
         )
         trainer = RLOOTrainer(
             model=model_id,
-            reward_funcs="trl-internal-testing/tiny-Qwen2ForSequenceClassification-2.5",
+            reward_funcs=reward_func,
+            args=training_args,
+            train_dataset=dataset,
+        )
+
+        previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
+
+        trainer.train()
+
+        self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
+
+        for n, param in previous_trainable_params.items():
+            new_param = trainer.model.get_parameter(n)
+            self.assertFalse(torch.equal(param, new_param), f"Parameter {n} has not changed.")
+
+    @require_vision
+    def test_training_vlm_multi_image(self):
+        dataset = load_dataset("trl-internal-testing/zen-multi-image", "conversational_prompt_only", split="train")
+
+        # For now, mixing image+text and text-only examples is not supported, so we filter out text-only examples
+        dataset = dataset.filter(lambda x: len(x["images"]) > 0)
+
+        def reward_func(completions, **kwargs):
+            """Reward function that rewards longer completions."""
+            return [float(len(completion[0]["content"])) for completion in completions]
+
+        training_args = RLOOConfig(
+            output_dir=self.tmp_dir,
+            learning_rate=0.1,  # increase the learning rate to speed up the test
+            per_device_train_batch_size=3,  # reduce the batch size to reduce memory usage
+            num_generations=3,  # reduce the number of generations to reduce memory usage
+            max_completion_length=8,  # reduce the completion length to reduce memory usage
+            max_prompt_length=None,  # disable prompt truncation, because usually, models don't support it
+            report_to="none",
+        )
+        trainer = RLOOTrainer(
+            model="trl-internal-testing/tiny-Qwen2_5_VLForConditionalGeneration",
+            reward_funcs=reward_func,
             args=training_args,
             train_dataset=dataset,
         )