prefetch weights while waiting for pending requests to complete#728
Merged
Conversation
|
@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:
|
facebook-github-bot
pushed a commit
that referenced
this pull request
Jan 23, 2026
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
Contributor
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): |
Contributor
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 ?
Contributor
Author
There was a problem hiding this comment.
I'll simplify this. The rest of the test is quite essential.
| logger.info( | ||
| "[ForgeMonarchExecutor] Deserializing TorchStore Controller from environment..." | ||
| ) | ||
| self.torchstore_controller = cloudpickle.loads( |
| model: str | ||
| iterations: int | ||
| prefetch_enabled: bool | ||
| n_fetcher_procs: int |
Contributor
There was a problem hiding this comment.
Can we test how this parameter affects throughput?
Contributor
Author
There was a problem hiding this comment.
n_fetcher_procs -- too high will slightly degrade the overall time.
prefetch_enabled -- this is the major toggle.
Contributor
There was a problem hiding this comment.
Right but is there a graph of where "too high" is? I imagine that also too low will not be optimized.
JenniferWang
added a commit
that referenced
this pull request
Jan 26, 2026
Summary: Pull Request resolved: #728 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
JenniferWang
force-pushed
the
export-D91092833
branch
from
969b8ab to
bfdb049
Compare
facebook-github-bot
pushed a commit
that referenced
this pull request
Jan 26, 2026
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
bfdb049 to
3f869a3
Compare
JenniferWang
added a commit
that referenced
this pull request
Jan 26, 2026
Summary: Pull Request resolved: #728 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
JenniferWang
force-pushed
the
export-D91092833
branch
from
3f869a3 to
3a769bb
Compare
facebook-github-bot
pushed a commit
that referenced
this pull request
Jan 27, 2026
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
3a769bb to
327d32a
Compare
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
pushed a commit
that referenced
this pull request
Jan 27, 2026
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
2 times, most recently
from
6f9290d to
f5fe7d6
Compare
allenwang28
approved these changes
Jan 27, 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.
joecummings
approved these changes
Jan 27, 2026
2 tasks
HosseinKaviani-H
pushed a commit
to HosseinKaviani-H/forge
that referenced
this pull request
Feb 9, 2026
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
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