feat(refit): add ModelExpress Megatron-to-vLLM reshard path - #3632

Open

KavinKrishnan wants to merge 25 commits into

NVIDIA-NeMo:mainfrom

KavinKrishnan:kavink/upstream-mx-megatron-publisher

Commits on Aug 19, 2026

feat(refit): classify native Megatron parameters for reshard publishing
KavinKrishnan
committed
1f02045
feat(refit): translate Megatron shards for ModelExpress publishing
KavinKrishnan
committed
c0c6f40
refactor(refit): narrow the Megatron adapter to the main-native contract
KavinKrishnan
committed
00f648a
feat(refit): add the mx_reshard weight-sync transport
KavinKrishnan
committed
f6b56b2
fix(grpo): shut down cleanly when a run raises
KavinKrishnan
committed
43b19dc
test(refit): add an mx_reshard GRPO functional test and bench configs
KavinKrishnan
committed
4a3eb10
perf(refit): report the receiver's quorum cost separately
KavinKrishnan
committed
35143fc
fix(refit): never let phase telemetry fail a refit
KavinKrishnan
committed
23c5713
perf(refit): ask the quorum only for the version stamp
KavinKrishnan
committed
f52d8c3
test(refit): add a larger-sample config for the MoE KL question
KavinKrishnan
committed
7ee91a9
feat(refit): add parameter-equality verification for mx_reshard
KavinKrishnan
committed
ebb3b8d
fix(tests): let the mx_reshard functional gate actually launch
KavinKrishnan
committed
74c0d2d
test(refit): stop the mx_reshard gate from going vacuous
KavinKrishnan
committed
c59cd8d
test(refit): record the FP8 arm's required override and its known failure
KavinKrishnan
committed
b040ca8
test(refit): add a 12-step correctness arm for the long horizon
KavinKrishnan
committed
111149c
test(refit): add the KL-floor arm that decomposes gen_kl_error
KavinKrishnan
committed
1bf4276
test(refit): gate on js_divergence_error instead of an absolute KL bound
KavinKrishnan
committed
9a2865d
test(bench): add variance attribution for run-to-run refit spread
KavinKrishnan
committed
0e33793
fix(refit): publish per-layer global QKV geometry
KavinKrishnan
committed
3aedf6c
chore(refit): keep private cluster configs out of the PR
KavinKrishnan
committed
f8eadfb
test(refit): cover MX handshake selection after main rebase
KavinKrishnan
committed
4716646
docs(refit): scrub internal references from reshard recipes
KavinKrishnan
committed
3582599
style(refit): satisfy ruff format across the mx_reshard sources
KavinKrishnan
committed
8e0afba
fix(refit): pick the MX discovery call by signature, not by TypeError
KavinKrishnan
committed
82fae40
fix(refit): close review gaps in the publish and verify paths
KavinKrishnan
committed
b28a2fd

Navigation Menu

Uh oh!

feat(refit): add ModelExpress Megatron-to-vLLM reshard path - #3632

feat(refit): add ModelExpress Megatron-to-vLLM reshard path#3632
KavinKrishnan wants to merge 25 commits into
NVIDIA-NeMo:mainfrom
KavinKrishnan:kavink/upstream-mx-megatron-publisher

Commits on Aug 19, 2026

feat(refit): classify native Megatron parameters for reshard publishing

feat(refit): translate Megatron shards for ModelExpress publishing

refactor(refit): narrow the Megatron adapter to the main-native contract

feat(refit): add the mx_reshard weight-sync transport

fix(grpo): shut down cleanly when a run raises

test(refit): add an mx_reshard GRPO functional test and bench configs

perf(refit): report the receiver's quorum cost separately

fix(refit): never let phase telemetry fail a refit

perf(refit): ask the quorum only for the version stamp

test(refit): add a larger-sample config for the MoE KL question

feat(refit): add parameter-equality verification for mx_reshard

fix(tests): let the mx_reshard functional gate actually launch

test(refit): stop the mx_reshard gate from going vacuous

test(refit): record the FP8 arm's required override and its known failure

test(refit): add a 12-step correctness arm for the long horizon

test(refit): add the KL-floor arm that decomposes gen_kl_error

test(refit): gate on js_divergence_error instead of an absolute KL bound

test(bench): add variance attribution for run-to-run refit spread

fix(refit): publish per-layer global QKV geometry

chore(refit): keep private cluster configs out of the PR

test(refit): cover MX handshake selection after main rebase

docs(refit): scrub internal references from reshard recipes

style(refit): satisfy ruff format across the mx_reshard sources

fix(refit): pick the MX discovery call by signature, not by TypeError

fix(refit): close review gaps in the publish and verify paths