fix(qwen3_5): preserve packed-sample boundaries in GatedDeltaNet

[HuiyingLi]

Summary

Fixes #2131 — Qwen3_5GatedDeltaNet was leaking recurrent state across packed-sample boundaries because:

1. FLA chunk_gated_delta_rule was called without cu_seqlens → recurrent state ran continuously across the whole pack.
2. causal_conv1d_fn (4-tap depthwise conv, upstream of FLA) was called without seq_idx → sample-1's last 3 tokens bled into sample-2's first 3 tokens.

Layer-output diff between standalone and packed-then-sliced on a small reproducer:
4.011e-01 → 0.000e+00.

Mechanism

The fix rides on the existing patch_hf_model class-swap machinery that already wraps every Qwen3_5GatedDeltaNet instance into CPAwareGatedDeltaNet.

• Dense path (Qwen3_5ForCausalLM / Qwen3_5ForConditionalGeneration): a new Qwen3_5DecoderLayerWithPacking(Qwen3_5DecoderLayer) subclass overrides forward to derive cu_seqlens and indices from the indexed attention_mask and pass them to linear_attn as keyword arguments. patch_hf_model is extended to class-swap every Qwen3_5DecoderLayer instance to this subclass.
• MoE path (Qwen3_5MoeForConditionalGeneration): Qwen3_5MoeBlock(Block) already custom-built; its forward is overridden in place to do the same kwarg-threading. No class-swap needed.
• CPAwareGatedDeltaNet._forward_no_cp accepts the new kwargs. When packed: skips apply_mask_to_padding_states (the unpad does it stronger), unpads [B, T, H] → [1, total_valid, H] if there's actual padding (B>1 case), threads seq_idx into causal_conv1d_fn, threads cu_seqlens into chunk_gated_delta_rule, repads on exit. When not packed: original kwargs path verbatim — bit-for-bit unchanged.

Files changed

```
M nemo_automodel/_transformers/infrastructure.py
M nemo_automodel/components/distributed/cp_utils.py
M nemo_automodel/components/models/common/packing.py
A nemo_automodel/components/models/qwen3_5/decoder_layer.py
M nemo_automodel/components/models/qwen3_5_moe/cp_linear_attn.py
M nemo_automodel/components/models/qwen3_5_moe/model.py
M tests/unit_tests/models/qwen3_5/test_cp_linear_attn_patch.py
```

What does NOT change

• The unpacked path (no indexed mask) is bit-for-bit identical to before.
• _forward_with_cp is not modified — packing+CP support is intentionally out of scope (no shipped Qwen3.5 recipe sets cp_size > 1).

Verification

Unit + reproducer evidence

Test	Coverage	Result
tests/unit_tests/models/qwen3_5/test_cp_linear_attn_patch.py	35 cases incl. new TestPackingHelpers	all pass
logs/reproduce_deltanet_packing_via_subclass.py	First-DeltaNet-layer comparison (literal issue procedure) for B=1 packed-no-padding AND B=2 packed-with-padding	all per-doc max_abs_diff = 0.000e+00
logs/reproduce_deltanet_packing_logits.py	4-layer dense stack + final norm + lm_head	all per-doc logits diff = 0.000e+00
logs/reproduce_deltanet_packing_logits_moe.py	4-layer MoE stack + final norm + lm_head	all per-doc logits diff = 0.000e+00

Production-recipe verification

Captured live during `automodel <qwen3_5_4b_neat_packing.yaml>`:

```
[INFO] Patched 24 GatedDeltaNet modules (cp=False) with FSDP-safe fp32 param wrapping.

step 0 chunk_gated_delta_rule:
q=(1, 1503, 32, 128) bf16 needs_unpad=False
cu_seqlens=[0, 371, 495, 816, 1074, 1198, 1503] n_docs=6 doc_lens=[371, 124, 321, 258, 124, 305]

step 1 chunk_gated_delta_rule:
q=(1, 2150, 32, 128) bf16 needs_unpad=False
cu_seqlens=[0, 1699, 2150] n_docs=2 doc_lens=[1699, 451]
```

End-to-end: production recipe → indexed mask → Qwen3_5DecoderLayerWithPacking.forward → CPAwareGatedDeltaNet._forward_no_cp → FLA chunk_gated_delta_rule is now receiving real per-document boundaries. With the bug, FLA would have seen cu_seqlens=None and run all 1503 tokens as one sequence.

🤖 Generated with Claude Code

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.