feat(rocm): enable bitwise Qwen3 Dense VIME alignment - #388

Merged

Flink-ddd merged 32 commits into

testfrom

rocm-vime-qwen3-bitwise

Sep 5, 2026

Commits on Aug 31, 2026

Add VIME Qwen3-8B 200-round consistency experiment
frank-2077
committed
577d51f
Show single-step ablation points in plots
frank-2077
committed
f12b154
Make GRPO explicit in VIME experiments
frank-2077
committed
7a93b1f
Enable activation recompute for long GRPO responses
frank-2077
committed
b4fc7e8
Keep deterministic collective staging IPC compatible
frank-2077
committed
64a41f3
Tune GRPO experiment for long math responses
frank-2077
committed
c4040cd
Cap GRPO responses below CP attention OOM threshold
frank-2077
committed
0818028
Run VIME consistency matrix with TP4 CP2
inaniloquentee
committed
9caf032
Keep colocated TP4 actor resident
inaniloquentee
committed
bcb5e69
Align Megatron TP4 attention reduction tree
inaniloquentee
committed
5403df6

Commits on Sep 1, 2026

Pin VIME Qwen3 production attention backend
frank-2077
committed
77c9d77
Use fused attention for TP4 CP2 production arms
frank-2077
committed
07aede0
fix-native-production-logp-route
frank-2077
committed
576741f
decouple-cuda-graph-from-logp-route
frank-2077
committed
d2173e8
docs: publish VIME TP4 CP2 consistency results
frank-2077
committed
09695b0
docs: add VIME TP4 CP2 reproduction runbook
frank-2077
committed
e2f6279
optimize TP4 rollout scheduling

frank-2077
authored and
inaniloquentee
committed
ea46f22
docs: clarify VIME mismatch comparison
inaniloquentee
committed
a463e77
perf: write deterministic GEMMs into collective staging
inaniloquentee
committed
f50c367
feat: configure reference-model comparison runs
inaniloquentee
committed
b3bfd2b

Commits on Sep 4, 2026

perf: optimize strict VIME TP4/CP2 execution
frank-2077
committed
6a7646c
docs: publish optimized G11 200-step comparison
frank-2077
committed
40db4d3
style: format strict VIME integration changes
frank-2077
committed
c663a71
feat: add TP4 VIME supplement suites
frank-2077
committed
acdf7ae
test: shorten VIME precision supplements to eight rounds
inaniloquentee
committed
0ae2c7b

Commits on Sep 5, 2026

Merge pull request #377 from RL-Align/vime-qwen3-8b-tp4-cp2-200
Flink-ddd
authored
ccb70e3
feat(rocm): add end-to-end attention ablation matrix

codex
authored and
inaniloquentee
committed
65f28d2
feat(rocm): enable bitwise Qwen3 VIME alignment
inaniloquentee
committed
ef18a08
refactor(rocm): isolate strict VIME platform paths
inaniloquentee
committed
0bc2085
refactor(rocm): move deterministic RoPE to ROCm
inaniloquentee
committed
999d666
feat(rocm): complete deterministic VIME runtime
inaniloquentee
committed
6152f16
Merge remote-tracking branch 'origin/test' into pr388-conflict-resolve
inaniloquentee
committed
872780c