prefetch weights while waiting for pending requests to complete #728
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 D91092833. |
meta-cla
bot
added
the
CLA Signed
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #728 +/- ##
==========================================
- Coverage 78.33% 68.73% -9.61%
==========================================
Files 36 42 +6
Lines 4209 4455 +246
==========================================
- Hits 3297 3062 -235
- Misses 912 1393 +481 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
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 Reviewed By: allenwang28 Differential Revision: D91038187
Summary: Feature parity with v0: allow prefetching weights while waiting for the pending requests to finish. ## Test Plan Introduced a benchmark that simulates the on-going requests with actual weight sync logic. Reference Group (V0) ``` ================================================================================ WEIGHT SYNC BENCHMARK RESULTS ================================================================================ Model: Qwen/Qwen3-8B Model size: 15.26 GB Iterations: 3 Prefetch enabled: False -------------------------------------------------------------------------------- Metric Time (s) Throughput (GB/s) -------------------------------------------------------------------------------- Avg push_weights 5.102 s 2.99 GB/s Avg update_weights 43.738 s 0.35 GB/s Avg total (push + update) 48.840 s ================================================================================ ================================================================================ WEIGHT SYNC BENCHMARK RESULTS ================================================================================ Model: Qwen/Qwen3-8B Model size: 15.26 GB Iterations: 3 Prefetch enabled: True Fetcher procs: 8 -------------------------------------------------------------------------------- Metric Time (s) Throughput (GB/s) -------------------------------------------------------------------------------- Avg push_weights 5.208 s 2.93 GB/s Avg update_weights 29.602 s 0.52 GB/s Avg total (push + update) 34.810 s ================================================================================ ``` Test Group (V1) ``` ================================================================================ WEIGHT SYNC BENCHMARK RESULTS ================================================================================ Model: Qwen/Qwen3-8B Model size: 15.26 GB Iterations: 3 Prefetch enabled: False -------------------------------------------------------------------------------- Metric Time (s) Throughput (GB/s) -------------------------------------------------------------------------------- Avg push_weights 5.070 s 3.01 GB/s Avg update_weights 39.974 s 0.38 GB/s Avg total (push + update) 45.044 s ================================================================================ ================================================================================ WEIGHT SYNC BENCHMARK RESULTS ================================================================================ Model: Qwen/Qwen3-8B Model size: 15.26 GB Iterations: 3 Prefetch enabled: True Fetcher procs: 8 -------------------------------------------------------------------------------- Metric Time (s) Throughput (GB/s) -------------------------------------------------------------------------------- Avg push_weights 5.055 s 3.02 GB/s Avg update_weights 28.730 s 0.53 GB/s Avg total (push + update) 33.784 s ================================================================================ ``` ## Next Steps [-] implement the prefetch logic & shared memory [-] Add metric similar to generator v0 [ ] Perf/Throughput testing compared to generator v0 Differential Revision: D91092833
facebook-github-bot
force-pushed
the
export-D91092833
branch
from
0c99d56 to
969b8ab
Compare
joecummings
reviewed
Jan 23, 2026
Member
joecummings
left a comment
joecummings
left a comment
There was a problem hiding this comment.
I can't tell if this is supposed to be a stacked diff? Contains more than just prefetch information.
| """TitanTrainer with weight modification capabilities for benchmarking.""" | ||
|
|
||
| @endpoint | ||
| async def modify_weights(self, scale: float = 1.1): |
Member
There was a problem hiding this comment.
super nit: do we need to parameterize this? Can't we just 1) assume it's a floating point and 2) arbitrarily add or scale by X ?
| logger.info( | ||
| "[ForgeMonarchExecutor] Deserializing TorchStore Controller from environment..." | ||
| ) | ||
| self.torchstore_controller = cloudpickle.loads( |
| model: str | ||
| iterations: int | ||
| prefetch_enabled: bool | ||
| n_fetcher_procs: int |
Member
There was a problem hiding this comment.
Can we test how this parameter affects throughput?
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:
Feature parity with v0: allow prefetching weights while waiting for the pending requests to finish.
Test Plan
Introduced a benchmark that simulates the on-going requests with actual weight sync logic.
Reference Group (V0)
Test Group (V1)
Next Steps
[-] implement the prefetch logic & shared memory
[-] Add metric similar to generator v0
[ ] Perf/Throughput testing compared to generator v0
Differential Revision: D91092833