Repository navigation
Pin the ports image to Miles 386ae3b07 and regenerate the golden samples - #256
Merged
Merged
Conversation
This was referenced Oct 14, 2026
michaellchung
force-pushed
the
s23-pr-14-per-island
branch
from
October 14, 2026 11:32
9422e10 to
e277bd3
Compare
Rebuild yeto-miles-ports on michaellchung/miles s23-sao-streaming-roles (386ae3b07: 56068bd56 plus the LoRA-hybrid full-parameter transport and the SAO streaming primitives), update the pins, lineage, Dockerfile ARG and build record, and regenerate the decoupling golden samples per hash-migration.md. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The P0-1.11 GPU run probe (s23-g2-fpprobe-20261014a) failed before Miles started: _run_ports re-derived a LoRA config hash for every island, and a full-parameter island has no LoRA rank (TypeError on lora_r=None). run_miles already binds the full-parameter identity (model config hash) on miles_args; _run_ports now forwards it and derives LoRA identity only for LoRA islands. A test parses the launcher's dense island command and drives _run_ports to run_ports_island. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Miles' set_default_megatron_args forces the distributed optimizer on for Adam regardless of argv. Under it every main_param is a flat slice of the owned range even at DP=1, so the full-parameter shard manifest refuses it as an incomplete FP32 master. The SAO source branch turned the optimizer off for any external policy sync; the fork port kept that only for the LoRA hybrid. Set use_distributed_optimizer=False as a runtime attribute in full-parameter mode (applied after Miles' defaults, before Megatron's validate_args) and refuse a parsed namespace that still has it on. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
michaellchung
force-pushed
the
s23-pr-15-image-pins
branch
from
October 14, 2026 11:33
e701f8e to
1e8dc49
Compare
michaellchung
marked this pull request as ready for review
October 14, 2026 11:33
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.
Summary
Rebuilds
yeto-miles-portson michaellchung/miless23-sao-streaming-roles(386ae3b07) and pins it. This closes the stack's expected test failures.Motivation
PRs 01-05 moved
MILES_NEXT_COMMIT(56068bd56, then the LoRA-hybrid transport and SAO streaming primitives) but the image tag and digest still named 64b591a4b until a rebuild. This image is the one used by the P0-1.11 GPU run.What changed
pins.py,entry.py: new image tag/digest and Miles lineage for 386ae3b07;docker/miles-ports/DockerfileARG..../2026-10-14-386ae3b-2fa8801/(build-record.json, check.txt, image-manifest.json).yeto-framework-decoupling/hash-migration.md(entry added there).tests/test_rl_ports_image.py: one expectation updated.Also in this PR: two full-parameter fixes (appended at the end of the chain)
Forward the full-parameter identity through the ports island entry:_run_portsused the LoRA hash for full-parameter runs; it now forwards the full-parameter identity (island_entry.py, test intest_rl_launcher_dense_full.py).Keep whole FP32 masters for full-parameter Miles actors: full-parameter actors turn off the distributed optimizer, with a validate guard and 3 tests (config.py,test_rl_miles_adapter_config.py).Behavior changes / compatibility
test_rl_ports_image(2), the 8test_decoupling_goldensamples, andtest_rl_backend_identity::test_golden_samples_keep_both_old_hashes_and_carry_the_identity(golden 9 + ports_image 2).Tests
Local CPU:
test_rl_ports_image.py,test_decoupling_golden.py,test_rl_backend_identity.py: 57 passed. Full suite with cargo: queued behind the GPU run lock; loglogs/pr-split/full-s23-pr-15-image-pins.log(fill in before opening the PR).Not verified on GPU
The image is the one the P0-1.11 GPU run uses (in progress at the time of writing); results will be linked here before merge. DeepSeek V4 paths in this image are not run on GPU.
Depends on
#255. Tree of this branch equals
s23-g3-fullparam(s23-final + 7f9daece + b8d4653f + 166ed64d).🤖 Generated with Claude Code