Repository navigation
Conversation
The magi cache key only hashed configs and traced source files. Changing DTensor placements (SimpleFSDP fully_shard -> replicate on a 1-rank mesh) leaves both unchanged, so the old artifact - still issuing the AllGathers - was silently reused. Add compute_graph_hash(graph.code + per-input dtype/placements/mesh shape) to the key.
6 tasks done
Collaborator
Author
|
Superseded by #89 (rebased the still-useful parts onto current main). Closing this PR. |
jiahy0825
pushed a commit
that referenced
this pull request
Oct 15, 2026
…split graphs (#89) * fix: cache key on graph+placements, AOT JIT fallback, skip FSDP overlap on split graphs Unmerged remnants of #88 / #86 / #84 / #79, rebased onto current main: - Cache key includes graph.code and DTensor placements. Changing SimpleFSDP fully_shard -> replicate used to reuse the old artifact (1-rank AllGathers served as DtoD copies). - AOT compile failures fall back to JIT so MAGI_COMPILE_AOT=1 can stay process-wide (VAE AOT, SimpleFSDP DiT JIT). - enable_fsdp no longer runs fullgraph overlap when disable_graph_split is False (VAE in a monolithic FSDP process). #79's restart-analysis half already landed as #75. - Timing bench dim 128->1024 so launch jitter does not flake max/min < 1.2 on B300/H100. * fix(cache): hash input placements only, not graph.code Restart-analysis retraces with different FX node names, so including graph.code in the magi cache key made the second pass miss the first-pass directory (test_assert_cache_hit_succeeds_with_restart_analysis / test_restart_analysis_cache on pt212). DTensor fully_shard vs replicate is still distinguished by the input placement signature. * fix(cache): hash only DTensor placements so restart-analysis still hits * fix(aot): do not swallow AssertionError/OOM when AOT falls back to JIT
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
1 participant
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.
Problem
The magi cache key (
MagiBackend._init_cache) only hashed the compile/inductor configs, the compiler manager and the traced source files. Changing DTensor placements alone (e.g. SimpleFSDPfully_shard->replicateon a 1-rank mesh, athena commit 2fb95f5b) leaves all of them unchanged, so the old artifact was silently reused. On 400B 8xH200 the stale graph kept issuing 1-rank AllGathers (served by NCCL as DtoD copies): 474 GB / 223 ms per request, fully exposed, until the cache dir was removed by hand.Fix
Add
compute_graph_hash(graph, example_inputs)to the key:graph.codeplus each input's dtype, DTensor placements and mesh shape.Tests
tests/feature_tests/cache/test_placement_cache_key.py(1 GPU, toy SimpleFSDP MLP):test_bug_reproduced_without_fix: with the graph hash disabled, areplicaterun after afully_shardrun in the same cache root compiles 0 times (reuses the stale artifact).test_bug_fixed_placement_change_recompiles: with the fix, the placement change recompiles, and rerunning the same placements still hits the cache.