feat(models): dense Qwen3.5 custom attention

[hubert-marek]

Summary

Part 2/7 of the split of #2485. Based on #2942 (feat/sft-renderer-only) — only for stack ordering; the code depends on main alone.

Implements dense Qwen3.5 (Qwen3_5ForCausalLM / Qwen3_5Model) on the PrimeRL attention modules, reusing the Qwen3.5-MoE building blocks (GatedDeltaNet, gated SDPA/flash attention, RMSNorm, rotary embedding) so ring/ulysses CP substitution patches dense and MoE consistently:

• New models/qwen3_5 package: decoder layers dispatch on layer_types (linear_attention / full_attention), flash varlen cu_seqlens from position_ids, identity state-dict conversion (dense HF layout matches PrimeRL layout).
• Shared normalize_qwen3_5_attn_implementation on the MoE module (eager→sdpa, kernels-community/vllm-flash-attn3→flash_attention_3), consumed by both models' attention factories and _check_and_adjust_attn_implementation.
• Registers Qwen3_5TextConfig/Qwen3_5ForCausalLM in AutoConfig and the custom causal-LM mapping; hooks dense Qwen3_5GatedFlashAttention into substitute_ring_attn/substitute_ulysses_attn.

Engagement note: hub Qwen3.5 checkpoints ship composite (VLM) configs, and composite→custom dispatch (_CUSTOM_VLM_MAPPING) lands in #2945 together with the state-dict remap that extracts the text tower — registering it here would crash get_model on any composite checkpoint. Standalone, this PR's class engages for text-config checkpoints and is covered by the dispatch/state-dict unit tests; the e2e custom-vs-hf parity and perf comparison run at the #2945 tip (same tip, same data pipeline, only model.impl differs), where the attention code exercised is exactly this PR's.

KL mismatch table

20 RL steps, hendrycks-math, batch 64, Qwen3.5-0.8B, impl=custom, vLLM 0.24 (run):

steps	mismatch KL
0–19 (all)	0.0003

Bar is < 0.015; measured 50× under it, flat across all steps.

E2E (W&B project pr2485)

Loss parity + perf, custom vs hf at identical config/data/seed (8×H100, FA3, seq 8192, batch 64, cat packing; loaded class verified in logs). Run at the #2945 tip where composite→custom dispatch exists:

model	impl	loss @ steps 7/8/9	tok/s	runs
0.8B	hf	0.8243 / 0.7225 / 0.8211	87.0k	j4364q4m
0.8B	custom	0.8243 / 0.7225 / 0.8210	80.1k	thga61ca
9B	hf	0.5182 / 0.4532 / 0.5256	29.8k	zzazx2av
9B	custom	0.5182 / 0.4532 / 0.5255	27.6k	oionqht8

Parity to 1e-4 at both scales. The custom impl carries a consistent ~7–8% dense-text throughput cost (profiling follow-up; likely the gated-attention epilogue / varlen bookkeeping) — its value is capability: native packed-boundary consumption, CP hooks (+4% over the old sync path at cp=2, see #2946), MoE EP, and the VLM path, none of which the hf impl supports.

Review fix: Removed the redundant Qwen3.5 varlen batch-size guard because this path always receives one packed row. The other automated review findings are addressed by later PRs in the stack or unreachable — see the inline replies.

• MoE-text custom regression run — v23zn2vu (10/10 steps)
• Integration coverage: sandwiched by passing feat(sft): renderer-only SFT tokenization + local data files #2942 leg and feat(trainer): seq_lens as the packed-sample boundary contract (SFT + RL) #2944 leg (no impl-specific integration tests exist); unit dispatch tests green; full suite ran at feat(rl): multimodal context parallelism #2948's tip

🤖 Generated with Claude Code

---

Note

Medium Risk
New training forward path for a flagship architecture (hybrid linear/full attention + custom flash varlen); mistakes would affect loss/CP correctness, but scope is additive with tests and shared MoE primitives.

Overview
Adds a dense Qwen3.5 custom path (Qwen3_5ForCausalLM / Qwen3_5Model) wired into AutoModelForCausalLMPrimeRL via Qwen3_5TextConfig, reusing MoE building blocks (GatedDeltaNet, gated SDPA/flash, RMSNorm, rotary) with dense MLP instead of experts.

Decoder layers follow layer_types (linear_attention vs full_attention). Flash paths derive varlen cu_seqlens from position_ids for packed training; KV cache is explicitly unsupported.

Shared normalize_qwen3_5_attn_implementation maps eager→sdpa and hub FA3 aliases→flash_attention_3, used by both dense and MoE factories plus _check_and_adjust_attn_implementation. Dense Qwen3_5GatedFlashAttention is hooked into ring and Ulysses CP substitution alongside MoE.

State-dict conversion is identity for dense HF layouts. Unit tests cover HF key parity on meta, attention dispatch/normalization, and ring patching.

^{Reviewed by Cursor Bugbot for commit fd38616. Bugbot is set up for automated code reviews on this repo. Configure here.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 05690b5b6b

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[mikasenghaas]

no strong opinions here, looks fine to me

[S1ro1]

Left a comment, else lgtm