[feat] Add Qwen3 MoE true-on-policy parity - #30

Open

maocheng23 wants to merge 40 commits into

feat/true_on_policy_qwen_moe

Commits on Feb 25, 2026

[1/8] fix: misc compatibility fixes for PyTorch and TE (#2)

authored and
guapisolo
committed
b4d401c
[2/8] feat: support partial checkpoint loading (#3)

authored and
guapisolo
committed
6c3fb8c

Commits on Feb 26, 2026

[3/8] feat: add post-attention and post-MLP layernorm support (#4)

yueming-yuan
authored and
guapisolo
committed
3f0fcac
[4/8] fix: MLA RoPE triton kernel head indexing and v_dim=0 support (#5)

authored and
guapisolo
committed
e82d9ee
[5/8] feat: support MTP training in RL (#6)

authored and
guapisolo
committed
3011052
[6/8] feat: support rollout routing replay (R3) and bypass for MTP layers (#10)

guapisolo
and
yueming-yuan
committed
228a364
[7/8] feat: add INT4 fake QAT for MoE grouped linear (#9)

authored and
guapisolo
committed
a95ad14

Commits on Feb 28, 2026

[8/8] fix: CUDA IPC incompatibility from Megatron bump (#11)

guapisolo
and
claude
committed
57784b8
fix: dp_reshardable checkpoint backward compat in Megatron core

guapisolo
and
claude
committed
07220d2

Commits on Mar 4, 2026

Upgrade Megatron from Dec 17 to Feb 13

guapisolo
and
yueming-yuan
committed
038e8e5

Commits on Mar 11, 2026

Commits on Mar 13, 2026

[fix] fix mixed precision parameters load from checkpoint (#19)

yueming-yuan
and
fzyzcjy
authored
cf014f9
[feat] distributed optimizer low GPU memory resume from checkpoint (#20)

yueming-yuan
and
fzyzcjy
authored
90f69fa

Commits on Apr 14, 2026

[fix] Add detach() to fp32 param shard for leaf-tensor consistency
guapisolo
authored
32dbe9f

Commits on Apr 15, 2026

[fix] attention_output_gate TP slice when num_kv_heads < TP (#22)
Zhichenzzz
authored
923b85b

Commits on Apr 20, 2026

[fix] Enforce param dtype before wrap ddp (#24)
guapisolo
authored
d9a5080

Commits on May 18, 2026

[feat] Init true on policy with qwen_dense

authored
23924a0

Commits on May 23, 2026

Add Qwen3 MoE true-on-policy Megatron parity

committed
f7a520a
Fix MoE true-on-policy weight refresh path

committed
9fe48e6
Split true-on-policy MoE helpers

committed
4c04bb2
Extract true-on-policy MoE layer extensions and combine ordering

committed
97dd9c9
Improve MoE true-on-policy parity kernels

committed
ccadee7
Improve MoE true-on-policy training path
maocheng23
committed
995fad7
Remove MoE straight-through fallback
maocheng23
committed
e5ab72a
Drop GroupedMLP fallback dtype patch
maocheng23
committed
ef0fd3e
Prune unused MoE parity fallback paths
maocheng23
committed
6a88277
Make true-on-policy MoE hook policy driven
maocheng23
committed
ea3ce92
Clarify true-on-policy MoE dispatch gate
maocheng23
committed
666ad16
Align MoE direct predicate with true-on-policy contract
maocheng23
committed
24088bf
Simplify MoE true-on-policy mode gate
maocheng23
committed
cfe1c38
Drive MoE kernel gate from true-on-policy contract
maocheng23
committed
c32f169
Align MoE contract gate with direct SGLang path
maocheng23
committed
a143c75
Remove true-on-policy debug dump plumbing
maocheng23
committed
dcf88ac
Keep GRPO loss unchanged for MoE parity
maocheng23
committed
5954723
Remove batch-invariant escape hatch
maocheng23
committed
1ef3a0e
Share SGLang MoE forward path
maocheng23
committed
ae93354
Simplify MoE: one forward, remove defensive bloat

maocheng23
and
cursoragent
committed
248d04b
Fix SGLang MoE weight adapter init
maocheng23
committed
1e562d7
Update SGLang fused MoE imports
maocheng23
committed
b0679e2