Repository navigation
Fail only the request whose forward raised - #981
Merged
Merged
Conversation
Gaurav-Shah05
left a comment
Gaurav-Shah05
left a comment
Collaborator
There was a problem hiding this comment.
LGTM. I checked the TP case against main. A forward that raises on every rank fails the same row on all and thus keeps the ranks in lockstep. But, if it's only raised on only some ranks, then it puts them out of step on main too.
On the unbatched path each request runs its own plan, forward and commit, but a raise in one of them left the loop and failed the whole batch, and with it everything else in flight on the worker. The loop now records the error on that rid, which the worker drops and fails as it does a prepare_inputs failure, and runs the rest. The slot fence is recorded on every exit of a rid, so a later rid on the same slot still waits out a raised one's queued copy, and the loop's own way out always drains next_slot before it releases commit_done. Postprocess skips a rid whose forward raised, so the client gets the forward's error. A raised rid keeps its admitted pages and in-flight mark until the removal its failure leads to, as a batch that raises does: unwinding them on the GPU thread would free pages the forward's queued kernels may still write.
No behaviour change. Drops the two-rank gloo test (both ranks run the same fake, so it can't catch an asymmetric failure), the removal test that repeats the real-resource test's tail, and the forward case the real-resource test covers; folds the every-rid-raises case into the parametrized one.
vasilevklart
force-pushed
the
users/kirill/per-rid-isolation
branch
from
November 1, 2026 22:54
4276d85 to
8512db1
Compare
dghosef
added a commit
that referenced
this pull request
Nov 1, 2026
Main brings #981 (per-rid forward failures), #985, #367 and #916 (chunked prefill as a graph Loop), #988 (Qwen3.5 on Blackwell), #977, #857 (stop policies checked in the graph runtime), #852, #1076, #943, #1077 and #938. Resolutions: - Piecewise leases: main's #977 design. A forward replays a lease_before_step region exactly when its label is in ModelInputsFromEngine.piecewise_leases; #859's engine-side filter (SubmoduleManagement.runners_for) and its tests go. The forwards #859 had moved onto runner presence check the lease again: Omnilingual's encoder loop and OmniVoice's backbone (both new here) as main's do, and V-JEPA2 and the FastConformer stream take main's code as it is. - Qwen3.5: main's token_stop (runtime_check) beside #859's DeviceLoopback; main's _splice_row_ids replaces both loop-back paths in preprocess; decode builds its rotary tables in the graph always (#859), text prefill and mixed steps under ingraph_decode_rope (#988), whose knob entry, docstring and docs now say so; main's compile_deterministic and speculative_input_seq_len. - TokenStop carries #859's withhold and #857's runtime_check. - Orpheus: #859's TokenStop with runtime_check, so the runtime derives its EOS policy (the loop cap is the budget, as its check_stop has it), in place of #857's hand-written AR policy; test_stop_policies passes its TokenStop. - Engine: chunking imports beside cudnn_benchmark. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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
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.
On main, a raise in one request's forward inside
Engine._exec_per_request, which runs unbatchable walks, fails every request in flight on the worker.Now only the raising rid fails;
_exec_single, for one-request or captured steps, still fails the batch. The slot fence andcommit_donerun on every exit, postprocess skips a failed rid so its first error is the one reported, and a failed rid keeps its pages until its REMOVE, since its kernels may still write them. Under TP/SP the ranks stay in step only if every rank raises on the same rid; if only some do, it breaks as on main.bagel_single_gpu.yaml), a hook raises after one request's forward; 120 s at c=8 and c=16 with cancelsMeasured on
6e78192ea(main237832f8a) before the rebase; the clean batch's images are bit-identical across arms.Tests: 6 cases in
test_admit_failure_handling.pyandtest_admit_rollback.pyfail on main and pass here; withtest_worker_admit_lifecycle.py, 53 pass on4276d852e. Onc313af056(mainf83fcaeb3), CPU Core fails 18 here and 17 on main, the difference being TP read-failure sims that time out under that load and the eval runner's SIGTERM test, each passing alone 3 of 3 on both; TP lockstep sim 11/11, ruff clean.Not run: a CPU-overhead harness; the batched path doesn't change, and the per-request path only adds a
try.Split out of #275 (threads 2, 7, 11); independent of #980.
AI assistance (Claude) was used for this PR.