Merge branch 'main' into weight-loading

Author: casteryh

.github/workflows/unit_test.yaml

@@ -32,6 +32,10 @@ jobs:
           eval "$(ssh-agent -s)"
           ssh-add - <<< '${{ secrets.FORGE_GITHUB_CI_FOR_TORCHSTORE }}'
           python -m pip install git+ssh://[email protected]/meta-pytorch/torchstore.git
+      - name: Install torchtitan
+        run: |
+          pip install --pre torchtitan==0.1.0.dev20250826+cpu --extra-index-url https://download.pytorch.org/whl/nightly/cpu
+          pip install tyro
       - name: Install dependencies
         run: python -m pip install --no-build-isolation -e ".[dev]"
       - name: Run unit tests with coverage

.gitignore

@@ -193,3 +193,6 @@ cover/
 wandb/
 
 assets/wheels/vllm*.whl
+
+# DCP artifacts
+model_state_dict/

apps/grpo/main.py

@@ -7,7 +7,6 @@
 # Usage: python -m apps.grpo.main --config apps/grpo/qwen3_1_7b.yaml
 
 import asyncio
-import time
 import uuid
 from dataclasses import dataclass
 from typing import Any, Callable
@@ -17,39 +16,20 @@
 import torchstore as ts
 from datasets import load_dataset
 from forge.actors.policy import Policy
-from forge.actors.reference_model import ReferenceModel  # noqa: F401
+from forge.actors.reference_model import ReferenceModel
 from forge.actors.replay_buffer import ReplayBuffer
 from forge.actors.torchstore_utils import get_param_key
-from forge.actors.trainer import _qwen3_hf_to_vllm
+from forge.actors.trainer import RLTrainer
 from forge.cli.config import parse
 from forge.controller.actor import ForgeActor
 from forge.controller.provisioner import shutdown
 from forge.data.rewards import MathReward, ThinkingReward
-from forge.losses.grpo_loss import SimpleGRPOLoss
 from forge.util.metric_logging import get_metric_logger
 from monarch.actor import endpoint
 from omegaconf import DictConfig
-from torchstore.state_dict_utils import DELIM
-from torchtitan.config.job_config import Model as TitanJobModelConfig
-from transformers import AutoModelForCausalLM
 from vllm.transformers_utils.tokenizer import get_tokenizer
 
 
-def compute_logprobs(
-    logits: torch.Tensor, input_ids: torch.Tensor, temperature: float = 1.0
-) -> torch.Tensor:
-    context_length = logits.shape[1] - input_ids.shape[1]
-
-    # Truncate request logits and drop last
-    logits = logits[:, context_length - 1 : -1]
-
-    # Compute logprobs
-    logprobs = torch.log_softmax(logits / temperature, dim=-1)
-    logprobs = torch.gather(logprobs, 2, input_ids.unsqueeze(-1)).squeeze(-1)
-
-    return logprobs
-
-
 @dataclass
 class Episode:
     # TODO: add adtional layer for multi-turn
@@ -118,64 +98,64 @@ def new_group(
         return cls(str(group_id), episodes)
 
 
-@dataclass
-class Trainer(ForgeActor):
-    """GRPO Trainer implementation for policy optimization."""
-
-    model_name: str
-    learning_rate: float = 1e-5
-    beta: float = 0.1
-    device: torch.device | None = None
-    state_dict_key: str = "model_state_dict"
-    dp_rank: int = 0  # TODO: support data parallelism, hard code it for now
-
-    @endpoint
-    async def setup(self):
-        if self.device is None:
-            self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-
-        self.model = AutoModelForCausalLM.from_pretrained(
-            self.model_name,
-            dtype=torch.bfloat16,
-            trust_remote_code=True,
-        ).to(self.device)
-        self.model.train()
-
-        self.optimizer = torch.optim.AdamW(
-            self.model.parameters(), lr=self.learning_rate
-        )
-        self.optimizer.zero_grad()
+def collate(batches: list[list[Episode]]):
+    inputs = []
+    targets = []
+    for batch in batches:
+        request = [e.request_tensor for e in batch]
+        request = torch.stack(request)  # [b x s]
 
-        self.loss = SimpleGRPOLoss(self.beta)
+        response = [e.response_tensor for e in batch]
+        response = torch.stack(response)  # [b x s]
 
-        self.logger.info(f"Trainer model initialized on {self.device}")
+        ref_logprobs = [e.ref_logprobs for e in batch]
+        ref_logprobs = torch.stack(ref_logprobs).squeeze()  # [b x s]
 
-    @endpoint
-    async def train_step(self, batch: list[list[Episode]]):
-        microbatch = batch[self.dp_rank]
-        pad_id = microbatch[0].pad_id
+        advantages = [e.advantage for e in batch]
+        advantages = torch.tensor(advantages).unsqueeze(-1)  # [b x 1]
 
-        # prepare batch
-        request = [e.request_tensor for e in microbatch]
-        request = torch.stack(request).to(self.device)  # [b x s]
+        pad_id = batch[0].pad_id
+        mask = response != pad_id
 
-        response = [e.response_tensor for e in microbatch]
-        response = torch.stack(response).to(self.device)  # [b x s]
+        input = {"tokens": torch.cat([request, response], dim=1)}
+        target = {
+            "response": response,
+            "ref_logprobs": ref_logprobs,
+            "advantages": advantages,
+            "padding_mask": mask,
+        }
+        inputs.append(input)
+        targets.append(target)
+    return inputs, targets
 
-        ref_logprobs = [e.ref_logprobs for e in microbatch]
-        ref_logprobs = torch.stack(ref_logprobs).to(self.device).squeeze()  # [b x s]
 
-        advantages = [e.advantage for e in microbatch]
-        advantages = torch.tensor(advantages).to(self.device).unsqueeze(-1)  # [b x 1]
-        del batch
+def compute_logprobs(
+    logits: torch.Tensor, input_ids: torch.Tensor, temperature: float = 1.0
+) -> torch.Tensor:
+    context_length = logits.shape[1] - input_ids.shape[1]
+    logits = logits[:, context_length - 1 : -1]
+    logprobs = torch.log_softmax(logits / temperature, dim=-1).to(input_ids.device)
+    logprobs = torch.gather(logprobs, 2, input_ids.unsqueeze(-1)).squeeze(-1)
+    return logprobs
 
-        input_ids = torch.cat([request, response], dim=1)
-        mask = input_ids != pad_id
-        logits = self.model(input_ids=input_ids, attention_mask=mask).logits
-        logprobs = compute_logprobs(logits, response)
-        del logits
 
-        mask = response != pad_id
+def simple_grpo_loss(
+    logits: torch.Tensor,
+    response: torch.Tensor,
+    ref_logprobs: torch.Tensor,
+    advantages: torch.Tensor,
+    padding_mask: torch.Tensor,
+    beta: float = 0.1,
+) -> torch.Tensor:
+    logprobs = compute_logprobs(logits, response)
+    kl = torch.exp(ref_logprobs - logprobs) - (ref_logprobs - logprobs) - 1
+    per_token_policy_loss = torch.exp(logprobs - logprobs.detach()) * advantages
+    per_token_loss = -(per_token_policy_loss - beta * kl)
+    loss = (
+        ((per_token_loss * padding_mask).sum(dim=1))
+        / (padding_mask.sum(dim=1).clamp(min=1.0))
+    ).mean()
+    return loss
         loss = self.loss(logprobs, ref_logprobs, advantages, mask)
         loss.backward()
         self.optimizer.step()
@@ -223,38 +203,6 @@ async def compute(self, group: Group) -> list[float]:
         return advantages.squeeze(0).tolist()
 
 
-class RefModel(ForgeActor):
-    def __init__(self, model_name, device: torch.device | None = None):
-        super().__init__()
-        self.model_name = model_name
-
-        if device is None:
-            self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-        else:
-            self.device = device
-
-        self.model = AutoModelForCausalLM.from_pretrained(
-            model_name,
-            dtype=torch.bfloat16,
-            trust_remote_code=True,
-        ).to(self.device)
-        self.model.eval()
-
-        self.logger.info(f"Model initialized on {self.device}")
-
-    @endpoint
-    async def forward(self, episode: Episode) -> torch.Tensor:
-        req, res = episode.request_tensor, episode.response_tensor
-        input_ids = torch.cat([req, res]).to(self.device).unsqueeze(0)
-        mask = input_ids != episode.pad_id
-
-        with torch.inference_mode():
-            logits = self.model(input_ids=input_ids, attention_mask=mask).logits
-
-        input_ids = input_ids[:, len(req) :]
-        return compute_logprobs(logits, input_ids)
-
-
 @dataclass
 class DatasetActor(ForgeActor):
     """Actor wrapper for HuggingFace dataset to provide async interface."""
@@ -309,10 +257,7 @@ async def pad_token(self):
 
 async def main(cfg: DictConfig):
     """Main GRPO training loop with rollout and training processes."""
-    titan_model = TitanJobModelConfig(name="qwen3", flavor="1.7B")
-    # Get parameters from config with fallbacks
     group_size = cfg.group_size
-    model = cfg.model
     max_req_tokens = cfg.max_req_tokens
     max_res_tokens = cfg.max_res_tokens
     mlogger = get_metric_logger(
@@ -322,7 +267,7 @@ async def main(cfg: DictConfig):
     )
 
     # ---- Setup services ---- #
-    await ts.initialize()
+    await ts.initialize(strategy=ts.ControllerStorageVolumes())
     (
         dataloader,
         policy,
@@ -334,17 +279,18 @@ async def main(cfg: DictConfig):
     ) = await asyncio.gather(
         DatasetActor.options(**cfg.services.dataset).as_service(**cfg.dataset),
         Policy.options(**cfg.services.policy).as_service(**cfg.policy),
-        Trainer.options(**cfg.services.trainer).as_service(**cfg.trainer),
+        RLTrainer.options(**cfg.services.trainer).as_service(
+            **cfg.trainer, loss=simple_grpo_loss
+        ),
         ReplayBuffer.options(**cfg.services.replay_buffer).as_service(
-            **cfg.replay_buffer
+            **cfg.replay_buffer, collate=collate
         ),
         ComputeAdvantages.options(**cfg.services.compute_advantages).as_service(),
-        RefModel.options(**cfg.services.ref_model).as_service(**cfg.ref_model),
+        ReferenceModel.options(**cfg.services.ref_model).as_service(**cfg.ref_model),
         RewardActor.options(**cfg.services.reward_actor).as_service(
             reward_functions=[MathReward(), ThinkingReward()]
         ),
     )
-
     print("All services initialized successfully!")
 
     # ---- Core RL loops ---- #
@@ -358,6 +304,7 @@ async def continuous_rollouts():
                 return
             prompt, target = sample["request"], sample["target"]
             responses = await policy.generate.choose(prompt)
+            # TODO: this shall be part of the responses metadata instead of a separate call
             version = await policy.get_version.choose()
             group = Group.new_group(
                 group_id=rollout_count,
@@ -370,20 +317,36 @@ async def continuous_rollouts():
                 target=target,
             )
 
-            # TODO: Parallelize the following calculation
-            for episode, response in zip(group.episodes, responses.outputs):
-                episode.request_tokens = responses.prompt_token_ids
+            input_ids = torch.ones(
+                (group_size, max_req_tokens + max_req_tokens),
+                dtype=torch.long,
+                device="cuda",
+            )
+            # Populate episode info and calculate rewards
+            for i, (episode, response) in enumerate(zip(group.episodes, responses)):
+                episode.request_tokens = response.prompt_ids
                 episode.response_tokens = response.token_ids
                 episode.response = response.text
-                episode.ref_logprobs = await ref_model.forward.choose(episode)
+                input_ids[i, :max_req_tokens] = episode.request_tensor
+                input_ids[i, max_req_tokens:] = episode.response_tensor
                 episode.reward = await reward_actor.evaluate_response.choose(
                     prompt=prompt, response=response.text, target=target
                 )
+
+            # Calculate reference logprobs
+            ref_logits = await ref_model.forward.choose(input_ids)
+            ref_logprobs = compute_logprobs(ref_logits, input_ids[:, max_req_tokens:])
+            for i, episode in enumerate(group.episodes):
+                episode.ref_logprobs = ref_logprobs[i]
+            del ref_logits, ref_logprobs, input_ids
+
+            # Calculate advantages and add to replay buffer
             advantages = await compute_advantages.compute.choose(group)
             for episode, advantage in zip(group.episodes, advantages):
                 episode.advantage = advantage
                 await replay_buffer.add.choose(episode)
 
+            # Log metrics
             avg_response_len = (
                 sum(len(e.response_tokens) for e in group.episodes) / group_size
             )
@@ -402,7 +365,8 @@ async def continuous_training():
             if batch is None:
                 await asyncio.sleep(0.1)
             else:
-                loss = await trainer.train_step.choose(batch)
+                inputs, targets = batch
+                loss = await trainer.train_step.choose(inputs, targets)
                 training_step += 1
                 mlogger.log("loss/training_step", loss, training_step)
                 start_time = time.perf_counter()