Add Qwen3.5-4B math recipes and make Qwen3 dense recipes run on the transformers-5.8 / torch-2.11 stack

[iamziyuzhao]

Summary

This PR makes AstraFlow run both the Qwen3 dense math recipes and new
Qwen3.5 math recipes on the current transformers-5 / torch-2.11 stack.

It supersedes #11 (which added the Qwen3.5 recipe but left the existing dense
Qwen3 math recipes broken on transformers ≥ 5). Targets dev.

What changed

1. Make Qwen3 dense recipes run on transformers ≥ 5 (fsdp_engine.py, recipes, cli_args.py)

• The packed/varlen training forward now passes attention_mask=None for all
  archs. The old dict(full_attention=None, sliding_attention=None) form is a
  transformers-4.x relic: on transformers ≥ 5 a dense model (qwen3/qwen2) treats
  that dict as a precomputed mask, skips creation, and crashes. None is
  correct — masking is driven by cu_seqlens + position_ids. (Verified more
  correct than the old form even where it didn't crash: under SDPA the old dict
  path silently lost block-diagonal separation.)
• Dense recipes (qwen3-1.7b-m2po-2gpus-{full,delta}, qwen3-8b-m2po-{full,delta})
  and the cli_args.py default now set attn_impl: kernels-community/flash-attn2 —
  a prebuilt FlashAttention-2 kernel pulled from the HF kernels hub (build
  variant ABI-matched to torch 2.11+cu130). The literal flash_attention_2 is
  avoided because it loads the local flash-attn wheel, which is ABI-broken vs
  torch 2.11+cu130; sdpa/eager remain valid choices. (FA2-kernels and
  sdpa were confirmed eval-equivalent within single-sample noise; FA2 is the
  faster default.)

2. Add Qwen3.5-4B math recipes + enablement

• New examples/math/qwen3.5-4b-m2po-{full,delta}/ recipes (M2PO, ctx 8k,
  DeepScaleR, math_verify), with a README documenting the validated stack.
  Qwen3.5 is a hybrid Gated-DeltaNet + attention checkpoint; the trainer uses
  attn_impl: sdpa (GDN linear-attn via fla kernels, full-attn via sdpa) and
  inference uses attention_backend: flashinfer.
• model.py: register qwen3_5 in VALID_VISION_MODELS (the checkpoint ships
  as Qwen3_5ForConditionalGeneration, loaded via the ImageTextToText path;
  trained text-only).

Validation

Qwen3-1.7B dense, transformers 5.8.1: full + delta trained 300 steps, eval
@0/100/200/300. math500 avg@k: full 73.8 → 79.9, delta 72.4 → 80.2; overall
avg@k: full 31.3 → 38.8, delta 31.1 → 38.3. importance_weight ≈ 1.0 every step
(packed forward correct); zero crashes; resumed cleanly from an external kill.

Qwen3.5-4B (from #11): trains end-to-end and eval rises steadily.

Qwen3.5-4B (overall)	step 0	step 80	Δ
avg@k	47.8%	57.4%	+9.6
pass@k	56.5%	67.4%	+10.9

Qwen3.5-4B re-validated on the pinned 0.5.13.post1 release: both full
and delta re-run end-to-end on sglang==0.5.13.post1 — training completes,
full (shard_copy) and delta (~7× compressed) weight transfer both work, and
eval holds at baseline (~49–51% overall avg@k over a short run). No regression
vs the predecessor git build the table above came from.

Validated environment (Qwen3.5)

torch 2.11.0+cu130, transformers 5.8.1, kernels 0.14.1, SGLang 0.5.13.post1
(published release with qwen3_5 support) served via attention_backend: flashinfer, flash-linear-attention 0.5.0, flashinfer-python 0.6.12,
attn_impl=sdpa.

pyproject.toml pins the full validated stack: transformers==5.8.1 (with
kernels>=0.14,<0.15), torch==2.11.0, flash-linear-attention==0.5.0 (the
fast Qwen3.5 GDN kernels; optional, with a pure-torch fallback), and
sglang==0.5.13.post1 — the published release that ships qwen3_5 (the
older 0.5.12.post1 predated it), which pulls flashinfer 0.6.12 in automatically.
uv pip install -e ".[sglang]" resolves cleanly to exactly the validated
versions.