add tracer in v1 to log generator perf metrics #720
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@JenniferWang has exported this pull request. If you are a Meta employee, you can view the originating Diff in D91038187. |
meta-cla
bot
added
the
CLA Signed
2 tasks
Summary: ## Summary This diff introduces vLLM v1 integration for forge & Monarch that works for version > 0.13.0. Functionality wise, this diff implements: - Single-node TP (unoptimized, TCP-based proc communication) - Multi-node TP (same TCP mechanism) Pending work (next diff stack): First focus on Single-node TP - Unix socket-based communication (instead of TCP) - Weight sync integration - Logging integration After that, we can introduce Pipeline Parallelism: - Extend executor to capture stage graph (DAG-like execution pattern) ## Decisions 1: Integration Layer -- `AsyncLLM` We integrate at the AsyncLLM layer (https://blog.vllm.ai/2025/01/27/v1-alpha-release.html), which sits higher in the stack compared to our v0 approach that disassembled EngineCore and integrated at the Worker level. We pick this layer for these main reasons 1. Reduced maintenance cost: vLLM v1 refactored internals significantly (new EngineCore, Scheduler, KVCacheManager). Integrating at AsyncLLM isolates us from these changes -- we only need to implement the Executor interface, not patch internal scheduling or memory management. 2. Better fit for agentic RL: The offline LLM class batches requests synchronously via `llm.generate([prompts])`. AsyncLLM exposes an async generator interface (async for output in llm.generate(prompt)) that supports streaming, priority scheduling, and concurrent request handling required for online RL rollouts. ## Decision 2: Extension Points -- Executor + WorkerWrapperBase | **Class** | **Base Class** | **Location** | **Purpose** | |------------------------|--------------------------------------------------------|-------------------------|--------------------------------------------------------------------------------------| | MonarchExecutor | vllm.v1.executor.abstract.Executor | monarch_executor.py | Creates ProcMesh from HostMesh, spawns workers, manages collective_rpc() dispatch. | | WorkerWrapper | vllm.v1.worker.worker_base.WorkerWrapperBase + Actor | monarch_executor.py | Dual-inheritance wrapper exposing vLLM worker methods as Monarch endpoints. | | ForgeMonarchExecutor (next diff) | MonarchExecutor | forge_executor.py | Extends executor with TorchStore Controller handling for weight updates. | | ForgeWorkerWrapper (next diff) | WorkerWrapper | forge_executor.py | Extends worker with TorchStore weight loading capabilities. | | Generator | ForgeActor | generator.py | Forge-specific orchestration: provisions hosts, allocates GPUs, manages AsyncLLM. | **`MonarchExecutor` and `WorkerWrapper` are designed to be upstreamed to vLLM alongside the existing `RayDistributedExecutor`, enabling Monarch as a first-class distributed backend.** ## Decision 3: Executor Owns Workers Lifecycle The architecture aligns closer with vLLM's Ray executor pattern where: - **Caller (Generator) owns HostMesh**: Resource allocation (hosts, GPU IDs) - **Executor owns ProcMesh + Workers**: Execution lifecycle ``` ┌───────────────────────────────────────────────────────────────────────┐ │ Host Mesh │ │ │ │ ┌─────────────────────────────────────────────────────────────────┐ │ │ │ Caller process │ │ │ │ │ │ │ │ ┌─────────────────────┐ ┌─────────────────────────────┐ │ │ │ │ │ AsyncLLM │ │ WorkerRegistry (actor) │ │ │ │ │ └─────────────────────┘ └─────────────────────────────┘ │ │ │ │ │ │ │ │ │ │ serialize host_mesh & registry to env vars │ │ │ │ ▼ │ │ │ │ ┌───────────────────────────────────────────────────────────┐ │ │ │ │ │ EngineCore subprocess │ │ │ │ │ │ │ │ │ │ │ │ MonarchExecutor │ │ │ │ │ │ ├── deserialize host_mesh │ │ │ │ │ │ ├── create proc_mesh from host_mesh (owns lifecycle) ───│──│──│──┐ │ │ │ ├── spawn worker actors on proc_mesh │ │ │ │ │ │ │ └── register workers in WorkerRegistry │ │ │ │ │ │ └───────────────────────────────────────────────────────────┘ │ │ │ │ └─────────────────────────────────────────────────────────────────┘ │ │ │ │ │ │ ┌─────────────────────────────────────────────────────────────────┐ │ │ │ │ GPU ProcMesh (owned by MonarchExecutor) │ │ │ │ │ │ │ │ │ │ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │ │ │ │ │ │ Worker 0 │ │ Worker 1 │ │ Worker 2 │ │ Worker 3 │ ... ◀──│──│──┘ │ │ └──────────┘ └──────────┘ └──────────┘ └──────────┘ │ │ │ │ ◀──── NCCL (tensor parallel) ────▶ │ │ │ └─────────────────────────────────────────────────────────────────┘ │ └───────────────────────────────────────────────────────────────────────┘ ``` **Design**: Caller owns host_mesh (resource allocation), executor owns proc_mesh + workers (execution). This mirrors vLLM's Ray executor pattern. Since we want to collocate Generator Actor with the worker host mesh, it's easier to stick to caller owning host mesh **WorkerRegistry** bridges the process boundary -- MonarchExecutor (in subprocess) registers workers there, Generator queries it after AsyncLLM initialization. **Executor Cleanup Responsibility**: Since MonarchExecutor creates proc_mesh from host_mesh, it owns the cleanup: 1. `MonarchExecutor.shutdown()` destroys process groups on workers (prevents NCCL errors) 2. Stops proc_mesh 3. `Generator.shutdown()` only needs to stop generator_proc ## Limitations - **TP**: Supported (single-node and multi-node) - **PP**: NOT supported (would require DAG-like execution pattern) - Shared memory cache (`mm_processor_cache_type='shm'`) not supported - Symmetric memory all-reduce disabled (`VLLM_ALLREDUCE_USE_SYMM_MEM=0`) ## Test Plan [-] Resource / Lifecycle: `pytest tests/integration_tests/test_generator_lifecycle.py -v -s` [-] Single node TP local benchmark throughput test: `python -m benchmarks.generator.throughput --config apps/grpo/qwen3_1_7b.yaml benchmark.num_requests=10 benchmark.dataset=fixed benchmark.fixed_prompt="Tell me a joke" benchmark.num_samples=5` to verify the vllm instantiation on local host. [-] Single node TP MAST benchmark throughput test to verify vllm instantiation on remote host: https://www.internalfb.com/msl/studio/runs/mast/qwen3_1_7b_mast-eh7o6d%3APRODUCTION%3A0/logs?attempt=0&taskGroups=client%3A0&statusFilter=PENDING%2CRUNNING%2CCOMPLETE%2CFAILED%2CABANDONED%2CSTOPPING&logarithm=%7B%22after%22%3A10%2C%22before%22%3A20%7D [-] Multi-node (TP) MAST benchmark throughput test: https://www.internalfb.com/msl/studio/runs/mast/qwen3_1_7b_multinode_test-gr8aes%3APRODUCTION%3A0/logs?attempt=0&taskGroups=client%3A0&statusFilter=PENDING%2CRUNNING%2CCOMPLETE%2CFAILED%2CABANDONED%2CSTOPPING&logarithm=%7B%22after%22%3A10%2C%22before%22%3A20%7D Reviewed By: allenwang28 Differential Revision: D90280578
Summary: ## tl;dr Adds ForgeMonarchExecutor and ForgeWorkerWrapper to enable weight synchronization via TorchStore for RL training loops (e.g., GRPO). Specifically, the diff serialize the TochStore controller Actor to MonarchExecutor for sharing the controller. ## Test Plan [-] Weight update correctness test: `TORCHSTORE_RDMA_ENABLED=0 PYTHONPATH=. pytest -s tests/integration_tests/test_policy_update.py::TestWeightSync::test_sanity_check --config tests/integration_tests/fixtures/qwen3_1_7b_tp.yaml` [-] Local host: `python -m apps.grpo.main --config apps/grpo/qwen3_1_7b.yaml` [-] Remote host: https://www.internalfb.com/msl/studio/runs/mast/qwen3_1_7b_mast-cve6ce%3APRODUCTION%3A0/logs?attempt=0&taskGroups=trainer%3A0%2Cref_model_0%3A0%2Cgenerator_0%3A0%2Cclient%3A0&statusFilter=PENDING%2CRUNNING%2CCOMPLETE%2CFAILED%2CABANDONED%2CSTOPPING&logarithm=%7B%22after%22%3A10%2C%22before%22%3A20%7D ## Next Steps [ ] implement the prefetch logic & shared memory [ ] Add metric similar to generator v0 [ ] Perf/Throughput testing compared to generator v0 Differential Revision: D90775552
Summary: ## tl;dr Add tracer in v1 to log perf metrics to wandb ## V0 vs V1 Metrics Parity Comparison | Category | v0 Metric | v1 Metric | Parity | |----------|-----------|-----------|--------| | **Generate - Request Count** | `generator/generate/count_requests` (SUM) | `generator/generate/count_requests` (SUM) | ✅ Same | | **Generate - Completion Count** | `generator/generate/count_sequences_completed` (SUM) | `generator/generate/count_sequences_completed` (SUM) | ✅ Same | | **Generate - E2E Timing** | `generator_perf/generate/*` (Tracer, GPU) | `generator_perf/generate/*` (Tracer, GPU) | ✅ Same | | **Update - Pending Requests** | `generator_perf/update_weights/sum_pending_gen_requests` (SUM) | N/A - AsyncLLM handles internally |⚠️ Skip (by design) | | **Update - Wait for Generation** | `generator_perf/update_weights/avg_waiting_for_generation_duration_s` (MEAN) | `generator_perf/update_weights/pause_generation_duration_s` (MEAN) | ✅ Equivalent - renamed for clarity | | **Update - Fetch Weights** | `generator_perf/update_weights/wait_fetch_weights` (MEAN) | `generator_perf/update_weights/worker_load_weights_duration_s` (MEAN) | ✅ Equivalent - renamed for clarity | | **Worker - Update Timing** | `generator_perf/update_weights/generator_worker_update/*` (trace, GPU) | `generator_perf/update_weights/generator_worker_update/*` (trace, GPU) | ✅ Same | ## Test Plan Main GRPO app: `python -m apps.grpo.main --config apps/grpo/qwen3_1_7b.yaml` ``` wandb: Run `wandb offline` to turn off syncing. wandb: Syncing run drawn-waterfall-686 wandb: ⭐️ View project at https://meta.wandb.io/jiyue/grpo-training wandb: 🚀 View run at https://meta.wandb.io/jiyue/grpo-training/runs/6pltx38p wandb: Detected [openai] in use. .... rvability.metric_actors.GlobalLoggingActor global_logger>] === [global_reduce] - METRICS STEP 1 === ... generator/generate/count_requests: 13.0 generator/generate/count_sequences_completed: 96.0 generator_perf/generate/total_duration_avg_s: 3.6518315022786463 generator_perf/generate/total_duration_max_s: 9.2080615234375 generator_perf/update_weights/pause_generation_duration_s: 2.8634108749683946 generator_perf/update_weights/resume_generation_duration_s: 1.918897032737732e-05 generator_perf/update_weights/worker_load_weights_duration_s: 3.506648204056546 ... ``` Make sure integration tests that do not initialize the tracer still works `pytest tests/integration_tests/test_generator_lifecycle.py -v -s` ## Next Steps [ ] implement the prefetch logic & shared memory [-] Add metric similar to generator v0 [ ] Perf/Throughput testing compared to generator v0 Differential Revision: D91038187
facebook-github-bot
force-pushed
the
export-D91038187
branch
from
6ffde9c to
072695e
Compare
allenwang28
approved these changes
Jan 22, 2026
Contributor
allenwang28
left a comment
allenwang28
left a comment
There was a problem hiding this comment.
Review automatically exported from Phabricator review in Meta.
facebook-github-bot
pushed a commit
that referenced
this pull request
Jan 23, 2026
Summary: ## tl;dr Add tracer in v1 to log perf metrics to wandb ## V0 vs V1 Metrics Parity Comparison | Category | v0 Metric | v1 Metric | Parity | |----------|-----------|-----------|--------| | **Generate - Request Count** | `generator/generate/count_requests` (SUM) | `generator/generate/count_requests` (SUM) | ✅ Same | | **Generate - Completion Count** | `generator/generate/count_sequences_completed` (SUM) | `generator/generate/count_sequences_completed` (SUM) | ✅ Same | | **Generate - E2E Timing** | `generator_perf/generate/*` (Tracer, GPU) | `generator_perf/generate/*` (Tracer, GPU) | ✅ Same | | **Update - Pending Requests** | `generator_perf/update_weights/sum_pending_gen_requests` (SUM) | N/A - AsyncLLM handles internally |⚠️ Skip (by design) | | **Update - Wait for Generation** | `generator_perf/update_weights/avg_waiting_for_generation_duration_s` (MEAN) | `generator_perf/update_weights/pause_generation_duration_s` (MEAN) | ✅ Equivalent - renamed for clarity | | **Update - Fetch Weights** | `generator_perf/update_weights/wait_fetch_weights` (MEAN) | `generator_perf/update_weights/worker_load_weights_duration_s` (MEAN) | ✅ Equivalent - renamed for clarity | | **Worker - Update Timing** | `generator_perf/update_weights/generator_worker_update/*` (trace, GPU) | `generator_perf/update_weights/generator_worker_update/*` (trace, GPU) | ✅ Same | ## Test Plan Main GRPO app: `python -m apps.grpo.main --config apps/grpo/qwen3_1_7b.yaml` ``` wandb: Run `wandb offline` to turn off syncing. wandb: Syncing run drawn-waterfall-686 wandb: ⭐️ View project at https://meta.wandb.io/jiyue/grpo-training wandb: 🚀 View run at https://meta.wandb.io/jiyue/grpo-training/runs/6pltx38p wandb: Detected [openai] in use. .... rvability.metric_actors.GlobalLoggingActor global_logger>] === [global_reduce] - METRICS STEP 1 === ... generator/generate/count_requests: 13.0 generator/generate/count_sequences_completed: 96.0 generator_perf/generate/total_duration_avg_s: 3.6518315022786463 generator_perf/generate/total_duration_max_s: 9.2080615234375 generator_perf/update_weights/pause_generation_duration_s: 2.8634108749683946 generator_perf/update_weights/resume_generation_duration_s: 1.918897032737732e-05 generator_perf/update_weights/worker_load_weights_duration_s: 3.506648204056546 ... ``` Make sure integration tests that do not initialize the tracer still works `pytest tests/integration_tests/test_generator_lifecycle.py -v -s` ## Next Steps [ ] implement the prefetch logic & shared memory [-] Add metric similar to generator v0 [ ] Perf/Throughput testing compared to generator v0 Reviewed By: allenwang28 Differential Revision: D91038187
facebook-github-bot
pushed a commit
that referenced
this pull request
Jan 23, 2026
Summary: ## tl;dr Add tracer in v1 to log perf metrics to wandb ## V0 vs V1 Metrics Parity Comparison | Category | v0 Metric | v1 Metric | Parity | |----------|-----------|-----------|--------| | **Generate - Request Count** | `generator/generate/count_requests` (SUM) | `generator/generate/count_requests` (SUM) | ✅ Same | | **Generate - Completion Count** | `generator/generate/count_sequences_completed` (SUM) | `generator/generate/count_sequences_completed` (SUM) | ✅ Same | | **Generate - E2E Timing** | `generator_perf/generate/*` (Tracer, GPU) | `generator_perf/generate/*` (Tracer, GPU) | ✅ Same | | **Update - Pending Requests** | `generator_perf/update_weights/sum_pending_gen_requests` (SUM) | N/A - AsyncLLM handles internally |⚠️ Skip (by design) | | **Update - Wait for Generation** | `generator_perf/update_weights/avg_waiting_for_generation_duration_s` (MEAN) | `generator_perf/update_weights/pause_generation_duration_s` (MEAN) | ✅ Equivalent - renamed for clarity | | **Update - Fetch Weights** | `generator_perf/update_weights/wait_fetch_weights` (MEAN) | `generator_perf/update_weights/worker_load_weights_duration_s` (MEAN) | ✅ Equivalent - renamed for clarity | | **Worker - Update Timing** | `generator_perf/update_weights/generator_worker_update/*` (trace, GPU) | `generator_perf/update_weights/generator_worker_update/*` (trace, GPU) | ✅ Same | ## Test Plan Main GRPO app: `python -m apps.grpo.main --config apps/grpo/qwen3_1_7b.yaml` ``` wandb: Run `wandb offline` to turn off syncing. wandb: Syncing run drawn-waterfall-686 wandb: ⭐️ View project at https://meta.wandb.io/jiyue/grpo-training wandb: 🚀 View run at https://meta.wandb.io/jiyue/grpo-training/runs/6pltx38p wandb: Detected [openai] in use. .... rvability.metric_actors.GlobalLoggingActor global_logger>] === [global_reduce] - METRICS STEP 1 === ... generator/generate/count_requests: 13.0 generator/generate/count_sequences_completed: 96.0 generator_perf/generate/total_duration_avg_s: 3.6518315022786463 generator_perf/generate/total_duration_max_s: 9.2080615234375 generator_perf/update_weights/pause_generation_duration_s: 2.8634108749683946 generator_perf/update_weights/resume_generation_duration_s: 1.918897032737732e-05 generator_perf/update_weights/worker_load_weights_duration_s: 3.506648204056546 ... ``` Make sure integration tests that do not initialize the tracer still works `pytest tests/integration_tests/test_generator_lifecycle.py -v -s` ## Next Steps [ ] implement the prefetch logic & shared memory [-] Add metric similar to generator v0 [ ] Perf/Throughput testing compared to generator v0 Reviewed By: allenwang28 Differential Revision: D91038187
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary:
tl;dr
Add tracer in v1 to log perf metrics to wandb
V0 vs V1 Metrics Parity Comparison
generator/generate/count_requests(SUM)generator/generate/count_requests(SUM)generator/generate/count_sequences_completed(SUM)generator/generate/count_sequences_completed(SUM)generator_perf/generate/*(Tracer, GPU)generator_perf/generate/*(Tracer, GPU)generator_perf/update_weights/sum_pending_gen_requests(SUM)generator_perf/update_weights/avg_waiting_for_generation_duration_s(MEAN)generator_perf/update_weights/pause_generation_duration_s(MEAN)generator_perf/update_weights/wait_fetch_weights(MEAN)generator_perf/update_weights/worker_load_weights_duration_s(MEAN)generator_perf/update_weights/generator_worker_update/*(trace, GPU)generator_perf/update_weights/generator_worker_update/*(trace, GPU)Test Plan
Main GRPO app:
python -m apps.grpo.main --config apps/grpo/qwen3_1_7b.yamlMake sure integration tests that do not initialize the tracer still works
pytest tests/integration_tests/test_generator_lifecycle.py -v -sNext Steps
[ ] implement the prefetch logic & shared memory
[-] Add metric similar to generator v0
[ ] Perf/Throughput testing compared to generator v0
Differential Revision: D91038187