[RL] - MessageEnv, Rollout types, Rubric, Renderer

.github/workflows/integration_test_4gpu_rl.yaml

@@ -64,10 +64,11 @@ jobs:
         # Install uv for faster dependency resolution
         pip install uv
 
-        # 1. Install Monarch and TorchStore
+        # 1. Install Monarch, TorchStore, and Renderers
         uv pip install torchmonarch
         uv pip install --no-deps "git+https://github.com/meta-pytorch/torchstore.git@main"
         uv pip install pygtrie portpicker
+        uv pip install "git+https://github.com/PrimeIntellect-ai/renderers.git@main"
 
         # 2. Install batch-invariant ops
         uv pip install --no-deps "git+https://github.com/thinking-machines-lab/batch_invariant_ops.git@main"

.github/workflows/integration_test_8gpu_rl_h100.yaml

@@ -66,10 +66,11 @@ jobs:
         # Install uv for faster dependency resolution
         pip install uv
 
-        # 1. Install Monarch and TorchStore
+        # 1. Install Monarch, TorchStore, and Renderers
         uv pip install torchmonarch
         uv pip install --no-deps "git+https://github.com/meta-pytorch/torchstore.git@main"
         uv pip install pygtrie portpicker
+        uv pip install "git+https://github.com/PrimeIntellect-ai/renderers.git@main"
 
         # 2. Install batch-invariant ops
         uv pip install --no-deps "git+https://github.com/thinking-machines-lab/batch_invariant_ops.git@main"

torchtitan/experiments/rl/README.md

@@ -27,11 +27,12 @@ uv venv --python 3.12 titan-rl
 source titan-rl/bin/activate
 ```
 
-1. Install Monarch and TorchStore from main:
+1. Install Monarch, TorchStore, and Renderers from main:
 ```bash
 uv pip install torchmonarch
 uv pip install --no-deps "git+https://github.com/meta-pytorch/torchstore.git@main"
 uv pip install pygtrie portpicker
+uv pip install "git+https://github.com/PrimeIntellect-ai/renderers.git@main"
 ```
 
 2. Install Flash Attention 3 kernels:

torchtitan/experiments/rl/actors/generator.py

@@ -152,9 +152,6 @@ def get_vllm_compilation_config(
 class SamplingConfig:
     """Sampling parameters passed to vLLM's SamplingParams."""
 
-    n: int = 8
-    """Number of completions to generate per prompt (vLLM SamplingParams.n)."""
-
     temperature: float = 0.8
     """Sampling temperature. 0.0 = greedy, higher = more random."""
 
@@ -258,6 +255,7 @@ def __init__(
         compile_config: CompileConfig,
         max_num_seqs: int,
         output_dir: str,
+        stop_token_ids: list[int],
     ):
         init_logger()
         sl.init_structured_logger(
@@ -275,9 +273,13 @@ def __init__(
         # the upper bound for concurrent sequences, determines KV-cache
         # block allocation (and therefore GPU memory usage), and bounds
         # the CUDA graph capture sizes.  Always computed by the caller
-        # (RLTrainer) as num_prompts_per_step * sampling.n.
+        # (RLTrainer) as num_groups_per_rollout_batch * group_size.
         self._max_num_seqs = max_num_seqs
 
+        # Renderer role-boundary stop tokens (e.g. Qwen3 `<|im_end|>`), injected by the
+        # controller;
+        self._stop_token_ids = stop_token_ids
+
         # Register TorchTitan model + parser with vLLM
         registry_to_vllm(
             model_spec,
@@ -388,17 +390,19 @@ async def generate(
         self,
         tokenized_prompts: list[list[int]],
         *,
+        request_ids: list[str],
         sampling_config: SamplingConfig | None = None,
         metrics_prefix: str = "generator",
     ) -> tuple[list[Completion], list[m.Metric]]:
         """Generate completions and generator metrics for tokenized prompts.
 
         Takes ``tokenized_prompts`` as ``[num_prompts][prompt_tokens]``.
-        Returns completions as ``[num_prompts * n]`` plus generator metrics,
-        where ``n`` is the resolved ``SamplingConfig.n`` completions per prompt.
+        Returns completions in the same order as ``request_ids`` plus generator
+        metrics.
 
         Args:
             tokenized_prompts: Tokenized prompts shaped ``[num_prompts][prompt_tokens]``.
+            request_ids: One id per prompt echoed on each ``Completion.request_id``.
             sampling_config: Optional per-call override for the generator's
                 default SamplingConfig. ``seed`` always comes from
                 ``config.debug.seed`` (not part of SamplingConfig).
@@ -419,22 +423,32 @@ async def generate(
                 temperature=_sampling_config.temperature,
                 top_p=_sampling_config.top_p,
                 max_tokens=_sampling_config.max_tokens,
-                n=_sampling_config.n,
+                n=1,  # group_size pre-expands prompts; the RL loop always samples n=1
+                stop_token_ids=self._stop_token_ids or None,
                 seed=self.config.debug.seed,
                 logprobs=1,
                 output_kind=RequestOutputKind.FINAL_ONLY,
             )
 
+            if len(request_ids) != len(tokenized_prompts):
+                raise ValueError(
+                    f"got {len(request_ids)} request_ids for {len(tokenized_prompts)} prompts"
+                )
+            if len(set(request_ids)) != len(request_ids):
+                raise ValueError(f"request_ids must be unique; got {request_ids}")
+
             # render_cmpl is vLLM's input-pipeline entry.
             # The tokenize step is a no-op for already-tokenized prompts. The
             # lower-level alternative is vllm.inputs.tokens_input; we use the
             # high-level API to stay resilient to vLLM internal changes.
             engine_inputs = self._engine.renderer.render_cmpl(
                 [{"prompt_token_ids": ids} for ids in tokenized_prompts]
             )
-            for i, engine_input in enumerate(engine_inputs):
+            for engine_input, request_id in zip(
+                engine_inputs, request_ids, strict=True
+            ):
                 self._engine.add_request(
-                    request_id=str(i),
+                    request_id=request_id,
                     prompt=engine_input,
                     params=sampling_params,
                 )
@@ -444,15 +458,14 @@ async def generate(
                 while self._engine.has_unfinished_requests():
                     all_outputs.extend(self._engine.step())
 
-            # vLLM may return requests out of order; sort by the integer
-            # request_id we assigned so prompt_idx lines up with the input.
-            all_outputs.sort(key=lambda o: int(o.request_id))
+            # Return completions in caller input order; the controller maps positionally.
+            request_order = {request_id: i for i, request_id in enumerate(request_ids)}
+            all_outputs.sort(key=lambda output: request_order[output.request_id])
 
             completions: list[Completion] = []
             generation_metrics: list[m.Metric] = []
             output_token_counts: list[int] = []
             for output in all_outputs:
-                prompt_idx = int(output.request_id)
                 generation_metrics.extend(
                     _prepare_generation_request_metrics(output, prefix=metrics_prefix)
                 )
@@ -465,8 +478,7 @@ async def generate(
                     completions.append(
                         Completion(
                             policy_version=self.policy_version,
-                            prompt_idx=prompt_idx,
-                            text=sample.text,
+                            request_id=output.request_id,
                             token_ids=sample.token_ids,
                             token_logprobs=per_token_logprobs,
                             finish_reason=sample.finish_reason,

torchtitan/experiments/rl/actors/trainer.py

@@ -456,7 +456,7 @@ async def forward_backward(
     async def optim_step(self) -> OptimStepOutput:
         """Clip gradients, step optimizer + LR scheduler, return updated state."""
         # TODO: Accept optional optimizer params (e.g. learning rate)
-        # to allow controller-owned schedules (see Tinker API).
+        # to allow controller-owned schedules.
 
         # capture LR before step
         current_lrs = self.lr_schedulers.schedulers[0].get_last_lr()

torchtitan/experiments/rl/batcher.py

@@ -230,11 +230,11 @@ def _pack_episodes(self, episodes: list[Episode]) -> Iterator[dict]:
         def _iterate_samples() -> Iterator[dict]:
             for ep in episodes:
                 prompt_len = len(ep.prompt_token_ids)
-                response_len = len(ep.token_ids)
-                raw_ids = ep.prompt_token_ids + ep.token_ids
-                gen_lp = [0.0] * prompt_len + ep.token_logprobs
-                loss_mask = [False] * prompt_len + [True] * response_len
-                advantages = [0.0] * prompt_len + [ep.advantage] * response_len
+                completion_len = len(ep.completion_token_ids)
+                raw_ids = ep.prompt_token_ids + ep.completion_token_ids
+                gen_lp = [0.0] * prompt_len + ep.completion_logprobs
+                loss_mask = [False] * prompt_len + [True] * completion_len
+                advantages = [0.0] * prompt_len + [ep.advantage] * completion_len
                 sample = {
                     "input_ids": raw_ids[:-1],
                     "labels": raw_ids[1:],

torchtitan/experiments/rl/config_registry.py

@@ -26,9 +26,10 @@
 from torchtitan.experiments.rl.actors.generator import SamplingConfig, VLLMGenerator
 from torchtitan.experiments.rl.actors.trainer import PolicyTrainer
 from torchtitan.experiments.rl.batcher import BatchConfig, Batcher
+from torchtitan.experiments.rl.examples.sum_digits import SumDigitsRollouter
 from torchtitan.experiments.rl.grpo import GRPOLoss, RLTrainer
 from torchtitan.experiments.rl.observability.metrics import MetricsProcessor
-from torchtitan.experiments.rl.sum_digits import SumDigitsEnv
+from torchtitan.experiments.rl.renderer import RendererConfig
 from torchtitan.models.common.attention import FlexAttention
 from torchtitan.models.qwen3 import model_registry
 from torchtitan.protocols.model import ModelConfigConverter
@@ -74,13 +75,12 @@ def rl_grpo_qwen3_0_6b_varlen() -> RLTrainer.Config:
         model_spec=model_registry("0.6B", attn_backend="varlen"),
         hf_assets_path="torchtitan/experiments/rl/example_checkpoint/Qwen3-0.6B",
         num_steps=10,
-        num_prompts_per_step=5,
+        num_groups_per_rollout_batch=5,
         num_validation_samples=20,
         compile=CompileConfig(enable=True, backend="aot_eager"),
-        env=SumDigitsEnv.Config(seed=42, correctness_reward=1.0, format_reward=0.3),
-        validation_env=SumDigitsEnv.Config(
-            seed=99, correctness_reward=1.0, format_reward=0.3
-        ),
+        rollouter=SumDigitsRollouter.Config(),
+        group_size=group_size,
+        renderer=RendererConfig(name="qwen3", enable_thinking=True),
         metrics=MetricsProcessor.Config(enable_wandb=True),
         batcher=Batcher.Config(
             batch=BatchConfig(local_batch_size=2, global_batch_size=8, seq_len=2048),
@@ -115,37 +115,34 @@ def rl_grpo_qwen3_0_6b_varlen() -> RLTrainer.Config:
             ),
             checkpoint=CheckpointManager.Config(enable=False),
             sampling=SamplingConfig(
-                n=group_size,
                 temperature=0.8,
                 top_p=0.95,
-                max_tokens=100,
+                max_tokens=700,
             ),
         ),
     )
 
 
 def rl_grpo_qwen3_0_6b_flex() -> RLTrainer.Config:
     """GRPO training config for Qwen3-0.6B with flex attention (4 GPUs: 2 gen + 2 train)."""
-    spec = model_registry("0.6B", attn_backend="flex")
     group_size = 8
     return RLTrainer.Config(
-        model_spec=spec,
+        model_spec=model_registry("0.6B", attn_backend="flex"),
         hf_assets_path="torchtitan/experiments/rl/example_checkpoint/Qwen3-0.6B",
         num_steps=10,
-        num_prompts_per_step=5,
+        num_groups_per_rollout_batch=5,
         num_validation_samples=20,
         # TODO: add aot_eager compiling overall, today it doesn't work because
         # we are missing mechanism to scoop Flex region to plug in inductor backend support
-        env=SumDigitsEnv.Config(seed=42, correctness_reward=1.0, format_reward=0.3),
-        validation_env=SumDigitsEnv.Config(
-            seed=99, correctness_reward=1.0, format_reward=0.3
-        ),
+        rollouter=SumDigitsRollouter.Config(),
+        group_size=group_size,
+        renderer=RendererConfig(name="qwen3", enable_thinking=True),
         metrics=MetricsProcessor.Config(enable_wandb=True),
         batcher=Batcher.Config(
             batch=BatchConfig(local_batch_size=2, global_batch_size=8, seq_len=2048),
         ),
         trainer=PolicyTrainer.Config(
-            optimizer=OptimizersContainer.Config(lr=2e-6),
+            optimizer=default_adamw(lr=2e-6),
             lr_scheduler=LRSchedulersContainer.Config(
                 warmup_steps=2,
                 decay_type="linear",
@@ -174,7 +171,6 @@ def rl_grpo_qwen3_0_6b_flex() -> RLTrainer.Config:
             ),
             checkpoint=CheckpointManager.Config(enable=False),
             sampling=SamplingConfig(
-                n=group_size,
                 temperature=0.8,
                 top_p=0.95,
                 max_tokens=100,
@@ -217,13 +213,12 @@ def rl_grpo_qwen3_1_7b() -> RLTrainer.Config:
         model_spec=model_registry("1.7B", attn_backend="varlen"),
         hf_assets_path="torchtitan/experiments/rl/example_checkpoint/Qwen3-1.7B",
         num_steps=10,
-        num_prompts_per_step=5,
+        num_groups_per_rollout_batch=5,
         num_validation_samples=20,
         compile=CompileConfig(enable=True, backend="aot_eager"),
-        env=SumDigitsEnv.Config(seed=42, correctness_reward=1.0, format_reward=0.3),
-        validation_env=SumDigitsEnv.Config(
-            seed=99, correctness_reward=1.0, format_reward=0.3
-        ),
+        rollouter=SumDigitsRollouter.Config(),
+        group_size=group_size,
+        renderer=RendererConfig(name="qwen3", enable_thinking=True),
         metrics=MetricsProcessor.Config(enable_wandb=True),
         batcher=Batcher.Config(
             batch=BatchConfig(local_batch_size=2, global_batch_size=8, seq_len=2048),
@@ -259,10 +254,9 @@ def rl_grpo_qwen3_1_7b() -> RLTrainer.Config:
             ),
             checkpoint=CheckpointManager.Config(enable=False),
             sampling=SamplingConfig(
-                n=group_size,
                 temperature=0.8,
                 top_p=0.95,
-                max_tokens=100,
+                max_tokens=700,
             ),
         ),
     )
@@ -275,13 +269,12 @@ def rl_grpo_qwen3_14b() -> RLTrainer.Config:
         model_spec=model_registry("14B", attn_backend="varlen"),
         hf_assets_path="torchtitan/experiments/rl/example_checkpoint/Qwen3-14B",
         num_steps=10,
-        num_prompts_per_step=5,
+        num_groups_per_rollout_batch=5,
         num_validation_samples=20,
         compile=CompileConfig(enable=True, backend="aot_eager"),
-        env=SumDigitsEnv.Config(seed=42, correctness_reward=1.0, format_reward=0.3),
-        validation_env=SumDigitsEnv.Config(
-            seed=99, correctness_reward=1.0, format_reward=0.3
-        ),
+        rollouter=SumDigitsRollouter.Config(),
+        group_size=group_size,
+        renderer=RendererConfig(name="qwen3", enable_thinking=True),
         metrics=MetricsProcessor.Config(enable_wandb=True),
         batcher=Batcher.Config(
             batch=BatchConfig(local_batch_size=2, global_batch_size=8, seq_len=2048),
@@ -316,17 +309,16 @@ def rl_grpo_qwen3_14b() -> RLTrainer.Config:
             ),
             checkpoint=CheckpointManager.Config(enable=False),
             sampling=SamplingConfig(
-                n=group_size,
                 temperature=0.8,
                 top_p=0.95,
-                max_tokens=100,
+                max_tokens=700,
             ),
         ),
     )
 
 
 def rl_grpo_qwen3_0_6b_batch_invariant() -> RLTrainer.Config:
-    """On-policy GRPO config for Qwen3-0.6B under same parallelism (4 GPUs: 2 gen + 2 train).
+    """On-policy GRPO config for Qwen3-0.6B (4 GPUs: 2 gen + 2 train).
 
     Enables deterministic + batch-invariant mode for true on-policy RL training.
     """
@@ -336,13 +328,12 @@ def rl_grpo_qwen3_0_6b_batch_invariant() -> RLTrainer.Config:
         model_spec=model_registry("0.6B", attn_backend="varlen"),
         hf_assets_path="torchtitan/experiments/rl/example_checkpoint/Qwen3-0.6B",
         num_steps=10,
-        num_prompts_per_step=5,
+        num_groups_per_rollout_batch=5,
         num_validation_samples=20,
         compile=CompileConfig(enable=True, backend="aot_eager"),
-        env=SumDigitsEnv.Config(seed=42, correctness_reward=1.0, format_reward=0.3),
-        validation_env=SumDigitsEnv.Config(
-            seed=99, correctness_reward=1.0, format_reward=0.3
-        ),
+        rollouter=SumDigitsRollouter.Config(),
+        group_size=group_size,
+        renderer=RendererConfig(name="qwen3", enable_thinking=True),
         metrics=MetricsProcessor.Config(enable_wandb=True),
         batcher=Batcher.Config(
             batch=BatchConfig(local_batch_size=2, global_batch_size=8, seq_len=2048),
@@ -381,10 +372,9 @@ def rl_grpo_qwen3_0_6b_batch_invariant() -> RLTrainer.Config:
             ),
             checkpoint=CheckpointManager.Config(enable=False),
             sampling=SamplingConfig(
-                n=group_size,
                 temperature=0.8,
                 top_p=0.95,
-                max_tokens=100,
+                max_tokens=700,
             ),
             debug=batch_invariant_config,
         ),

torchtitan/experiments/rl/environment/__init__.py

@@ -0,0 +1,20 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+from torchtitan.experiments.rl.environment.message import (
+    MessageEnv,
+    MessageEnvInitOutput,
+    MessageEnvStepOutput,
+)
+from torchtitan.experiments.rl.environment.token import TokenEnv, TokenEnvOutput
+
+__all__ = [
+    "MessageEnv",
+    "MessageEnvInitOutput",
+    "MessageEnvStepOutput",
+    "TokenEnv",
+    "TokenEnvOutput",
+]