Repository navigation

feat: add experimental Qwen3.8-Flash-Next post-training support - #502

Draft

ruijieguo wants to merge 45 commits into

alibaba:mainfrom

ruijieguo:feat/qwen38-flash-next-training

Draft

feat: add experimental Qwen3.8-Flash-Next post-training support#502
ruijieguo wants to merge 45 commits into
alibaba:mainfrom
ruijieguo:feat/qwen38-flash-next-training

Commits on Sep 11, 2026
- Fix Qwen3.5 hybrid attention and add torch GDN fallback
  ruijieguo
  committed
  bd6335b
Commits on Sep 12, 2026
- chore: isolate unverified Qwen4 experimental adapter scaffold
  ruijieguo
  committed
  5a4b62a
- fix(qwen4-exp): apply exact GR residuals around raw transformer blocks
  ruijieguo
  committed
  9357c1e
- feat: add Qwen3.8 Flash-Next training path
  ruijieguo
  committed
  a320eae
- feat: integrate qwen38 flash next training lifecycle
  ruijieguo
  committed
  31af6e8
- fix: preserve Flash-Next assets on HF export
  ruijieguo
  committed
  f5781e1
- fix: write external asset sidecar on rank zero
  ruijieguo
  committed
  667f9d1
Commits on Sep 20, 2026
- feat: support Qwen3.8 Flash-Next post-training
  ruijieguo
  committed
  4d68571
- test: add bounded Qwen38 backbone payload validator
  ruijieguo
  committed
  0ebae8e
Commits on Sep 24, 2026
- fix: match Qwen38 native causal convolution product rounding
  ruijieguo
  committed
  a40838b
Commits on Sep 25, 2026
Commits on Sep 26, 2026
- fix: preserve short Flash-Next OPD rollouts
  ruijieguo
  committed
  c2c3fdb
- fix: guard Ray temporary storage for long runs
  ruijieguo
  committed
  fc34236
- fix: isolate Ray cluster when dedicated temp storage is configured
  ruijieguo
  committed
  5bcbc41
Commits on Sep 27, 2026
- fix: scope Ray cleanup and preserve rollout asset provenance
  ruijieguo
  committed
  84d005c
- fix: reject Ray temp paths that overflow Unix sockets
  ruijieguo
  committed
  a57d56a
- fix: honor explicit SFT pipeline max steps
  ruijieguo
  committed
  fb2e2ba
- test: enforce exact SFT pipeline step caps
  ruijieguo
  committed
  0959681
- test: prove SFT cap overrides epoch budget
  ruijieguo
  committed
  799a6a3
- fix: preserve epoch-based SFT defaults
  ruijieguo
  committed
  8f64df2
- fix: preserve native Flash-Next GR rounding during training
  ruijieguo
  committed
  7319152
- fix: retain checkpoint-native Flash-Next router precision
  ruijieguo
  committed
  383f069
- test: compare natural router precision using one model load
  ruijieguo
  committed
  bdecf47
- fix: preserve native GR projection rounding and adapter hooks
  ruijieguo
  committed
  e6de032
- fix: make Flash-Next LoRA shrink repeatable
  ruijieguo
  committed
  37e5756
Commits on Sep 28, 2026
- fix: scale QSA indexer distillation temperature to sqrt(head_dim)
  
  ruijieguo
  and
  claude
  committed
  8dcde5e
Commits on Oct 1, 2026
- Merge upstream v0.4.0 into feat/qwen38-flash-next-training
  
  ruijieguo
  and
  claude
  committed
  df1c3f2
- fix: adapt Qwen3.8 paths to upstream v0.4.0 interfaces after merge
  
  ruijieguo
  and
  claude
  committed
  d0098fe
- fix: keep CPU-master offload valid across park/restore on the backend store
  
  ruijieguo
  and
  claude
  committed
  436dcc6
- test: update stale fixtures for upstream v0.4.0 interfaces
  
  ruijieguo
  and
  claude
  committed
  1036281
- fix: keep grouped expert LoRA rank at r // moe_router_topk after upstream merge
  
  ruijieguo
  and
  claude
  committed
  d95dcb3
- fix: let vLLM dev builds pass the expert-parallel version gate by capability
  
  ruijieguo
  and
  claude
  committed
  9a72bc6
- fix: serialize TensorDict-backed columns in rollout dumps
  
  ruijieguo
  and
  claude
  committed
  d81b543
- fix: stop TrainingArguments from overriding Qwen4Exp's fused MoE permute
  
  ruijieguo
  and
  claude
  committed
  67c67e8
- test: make the EP dev-build test independent of torch.distributed state
  
  ruijieguo
  and
  claude
  committed
  73d95da
Commits on Oct 5, 2026
- test: give the qwen4_exp converter stubs the new use_value_head field
  
  ruijieguo
  and
  claude
  committed
  759463e
- fix: stop an undeclared reward dependency from breaking every RLVR run
  
  ruijieguo
  and
  claude
  committed
  4725516

Commits

Commits on Sep 11, 2026

Commits on Sep 12, 2026

Commits on Sep 20, 2026

Commits on Sep 24, 2026

Commits on Sep 25, 2026

Commits on Sep 26, 2026

Commits on Sep 27, 2026

Commits on Sep 28, 2026

Commits on Oct 1, 2026

Commits on Oct 5, 2026