Repository navigation
Conversation
When assert_cache_hit=True and a cache entry exists but load() returns None due to restart_analysis_count > 0, fall through to the normal compile path instead of raising RuntimeError. standalone_compile will trigger the same TensorifyScalarRestartAnalysis that occurred during bake, causing dynamo to re-trace with the correct graph shape. On retry, load() succeeds. Without this fix, verify_compile_cache.sh fails with a false cache miss for any subgraph that encountered RestartAnalysis during bake. Add regression test: test_assert_cache_hit_restart_analysis.py
When enable_fsdp is globally True but a model uses graph splitting (disable_graph_split=False), the FSDP fullgraph overlap pass would assert-fail. Add a disable_graph_split guard so non-FSDP models (VAE, etc.) compile correctly alongside FSDP models.
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Superseded by #89 (rebased the still-useful parts onto current main).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.
When
enable_fsdpis globallyTruebut a model uses graph splitting (disable_graph_split=False), the FSDP fullgraph overlap pass would assert-fail:This happens when compiling models like VAE alongside FSDP models (base, SR) in a monolithic deployment. The fix adds a
disable_graph_splitguard at the call site (line 672) so non-FSDP models compile correctly.Tested: 400B FSDP monolithic deployment on 8×H200 — all three compile phases (base, SR, VAE) completed successfully.