Add Qwen3.6-27B hybrid DeltaNet/GQA contrib model and vLLM serving path

[m-deepankar-singh]

Summary

This adds a Qwen3.6-27B contrib implementation for NxDI. Qwen3.6-27B is not a standard transformer-only decoder; it uses a hybrid [3 DeltaNet + 1 GQA] x 16 architecture with 48 recurrent DeltaNet/GDN layers and 16 GQA layers. Supporting it requires model code, DeltaNet NKI kernels, hybrid recurrent-state cache handling, vLLM/OpenAI serving glue, and validation coverage.

The main additions are:

• Qwen3.6/Qwen3.5-compatible model implementation under contrib/models/Qwen3.6-27B/.
• DeltaNet/GDN NKI kernels for token generation and chunked context encoding.
• Hybrid cache/APC support for matching attention KV reuse with GDN recurrent and convolution state checkpoints.
• Selective FP8 path for long-context serving while keeping sensitive recurrent state/cache paths in BF16/FP32.
• vLLM Neuron registry/launcher/proxy helpers and OpenAI-compatible validation scripts.
• Unit, integration, and artifact audit coverage for model config, weight conversion, Hybrid APC/cache semantics, scheduler behavior, and serving correctness.

Why this is larger than a normal contrib model

Stock NxDI has the standard serving substrate for transformer models, but Qwen3.6-27B needs additional hybrid recurrent-state semantics. Attention KV prefix reuse alone is insufficient. A safe reusable prefix is the intersection of attention KV cache hits, GDN recurrent checkpoint hits, and GDN convolution checkpoint hits. The PR therefore includes the minimal scheduler/model/cache bridge needed for coherent long-context Hybrid APC.

Validation

Recorded on trn2.3xlarge, TP=4, LNC=2. Evidence committed under validation_outputs/qwen36_nativechunk_baseline_20260609T000000Z/.

• 256K coherent native-chunk baseline, locally verified:
  • 16K native-chunk run: 16,374 prompt tokens, 6.84 s TTFT, 2,394.6 tok/s usage-accounted, pass=true, thinking enabled.
  • Long-context run: 242,864 usage.prompt_tokens, 235.98 s TTFT, 1,029.2 tok/s usage-accounted, pass=true, thinking enabled.
  • Empty invalid-token/fallback/NaN/NRT/traceback log scan; summary marks the run coherent and target-recovered.
• Correctness checks:
  • BF16 quality smoke: 7/7 prompts passed with enable_thinking=False.
  • HF greedy comparison: 156/160 token positions matched HF greedy (97.5%); 9/10 prompts matched exactly for all 16 generated tokens.
  • Strict Hybrid APC exactness passed for full-prefix, partial-prefix, and real-token generation cases.

Known limitations and follow-ups

• Continuous batching past max_num_seqs=1 is not included in this baseline.
• Some 32K/64K APC validation requires a recompiled artifact with larger prefix bucket coverage.
• Native Qwen MTP speculative decoding is not included.
• The stable direct-solve DeltaNet path trades some decode throughput for numerical stability versus the faster Neumann-style path.

Tests

• CPU unit tests for config parsing, weight conversion, Hybrid APC/cache behavior, DeltaNet decay, scheduler behavior, and artifact config audit.
• Integration test path (test/integration/test_model.py) for model load/generation/coherence on trn2.3xlarge.
• OpenAI/vLLM validation scripts for cold prefill, decode, APC exactness, boundary probes, and memory capture.

🤖 Generated with Claude Code