Conversation
- bind policy, model, router, checkpoint, token, and kernel semantics - encode compact IDs in a checksummed fail-closed safetensors artifact - add pure route-drift diagnostics and a codec-only benchmark - keep capture, transport, replay, admission, and training unchanged Benchmark: NVIDIA L20, PyTorch 2.11.0+cu128, codec-only synthetic 8192x40x22 trace, 512 experts, median of 50 after 5 warmups - ID storage: 57671680 -> 14417920 bytes (0.250x, -75.0%)
- freeze a per-response trace grid behind exact token and behavior identities - reject policy, ordering, padding, boundary, and cache mutations - preserve the existing rollout path when trace artifacts are absent - add deterministic binding and validation benchmarks Benchmark: NVIDIA L20, PyTorch 2.11.0+cu128, synthetic B=2 G=2 P=512 R=1024 L=40 K=22 E=512, median of 5 binds and 50 validations after 5 warmups - CPU IDs: 7.215 MB artifact, 83.732 ms/bind, 0.871 ms/validation - GPU-resident IDs+weights+margin: 14.753 MB artifact, 103.433 ms/bind, 1.206 ms/validation
- split bound rollout artifacts into deterministic authenticated frames - enforce byte, item, manifest, and total receive limits - reject corrupt, incomplete, reordered, and cross-batch shard sets - assign shards generically with real two- and four-rank Gloo coverage Benchmark: CPU, PyTorch 2.11.0+cu128, synthetic B=8 G=8 P=512 R=1024 L=40 K=22 E=512 IDs, median of 3 after 1 warmup - 115434944 payload -> 115472769 transport bytes (+0.0328%) - build: 244604 us median; verify all: 1153011 us median - reconstruction: 1143879 us median; 64 shards at 2 MiB cap
ViperEkura
force-pushed
the
main
branch
2 times, most recently
from
September 8, 2026 23:09
e7237f6 to
093c57e
Compare
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.
Summary
Depends on #63 (rollout identity), transitively #64 (RouteTraceV0).
Exact-head confirmation (2026-09-05)
python -m pytest -p no:cacheprovider -q tests/at953c7714ecb5ee3526193bf5fcf2112e74514777: 757 passed, 57 skipped in 63.65s, container exit 0.This final run retained its log independently of SSH connectivity. It used
the CPU-only environment described below; GPU tests remain skipped.
Add deterministic bounded byte shards for rollout-bound RouteTrace artifacts,
authenticate shards/items, validate assigned subsets, and reconstruct the exact
batch. Bound manifest/shard/item/total sizes and reject malformed, noncanonical,
corrupted, missing/reordered, cross-batch and trailing data.
Rank assignment is generic round-robin. Actual distributed tests use 2 or 4
ranks only. No network selection, model capture, replay, admission policy or
default training behavior changes are included.
Revision and dependency
e1e533c88e74efab4f379d6ccbc830c2b3e698e2.953c7714ecb5ee3526193bf5fcf2112e74514777.bea9ecf71f623d7b78aa2baed59646be871af19c(
codex/draft-rollout-route-identity), which depends on RouteTraceV0.Review the increment after those dependencies; main-based diff includes them.
Validation
python -m pytest -p no:cacheprovider -q tests/trainer/test_rollout_route_trace.py tests/trainer/test_rollout_route_transport_benchmark.py: 19 passed at7d54a9ac100341febf90c2bc5b4d1020005234e4, including CPU Gloo 2/4 ranks.import/format and
git diff --checkpassed.Initial collection lacked httpx2; a retry lost SSH before returning a final
result. Neither is counted as a successful full run.
CPU image
astrai-test:cu128, Python 3.12.3, torch 2.11.0+cu128, pytest 8.3.5.No GPUs exposed; httpx2 2.12.0 added only in the disposable test container.
Historical benchmark (pre-rebase)
CPU framing only, L20 host: IDs B=8 G=8 P=512 R=1024 L=40 K=22 E=512,
2 MiB / 8-item caps: 115,434,944 -> 115,472,769 bytes (+0.032767%), 64 shards;
build median 244,604 us, verify-all 1,153,011 us, reconstruct 1,143,879 us.
These numbers exclude actual networking and training throughput.
Review status
AI assistance was used. Keep Draft for human line-by-line review, dependency
review and CI; no fresh GPU or end-to-end performance claim is made.