[vivekkhandelwal1]

Dense Qwen3 models besides Qwen3.5, read from the model rather than hard-coded:

• Control-token ids and generation prompts come from the model's tokenizer and chat template (templates.ChatFormat): the rollout's stop token, the repair suffix check, the engines' self-test and the replies' stop stripping. A parse adapter no longer names the stop token.
• The trainer loads the base with its family's class (model_family.py, from config.json). The linear-attention fast-path check doesn't apply to a full-attention model. The LoRA targets, the merge tool and the slot writer accept Qwen3's model.layers.N names, including Qwen3-0.6B's single-file layout with a stored lm_head.weight.
• A CPU end-to-end test runs two wired iterations on a tiny Qwen3 (real control ids, its chat template, tied embeddings), with thinking off and on.

Measured on one MI350X GPU with the arithmetic example, through the unchanged run job:

• Qwen3-8B score (configs/arith-score-qwen3-8b.yaml): 256 of 256 scored, 226 correct, landing round trip PROVEN.
• Qwen3-8B, 3-iteration training: iteration 1 trained (127 of 128 samples) and its update LANDED, with every trainer check passing. The gap gate then stopped the run at iteration 2: all 8 samples of one prompt were 17 to 39 nats from the trainer's log-probs over 20-60% of their tokens, while the other 120 stayed under 0.5. That's systematic rather than numerical noise, and not explained yet; the gate did its job and nothing off-policy was trained on.
• Qwen3-0.6B: TokenSpeed at 3b684074 fails with an illegal memory access in its scheduler after two clean batches, in a score run and a training run alike, so its configs are replaced by Qwen3-8B ones.
• Qwen3.5-4B on this branch (regression check, configs/arith-score.yaml): 256 of 256 scored, 243 correct, landing PROVEN.

Stacked on #613. Draft until the iteration-2 gap is understood.

[Copilot]

🔵 Needs a closer look

Unexplained 17–39-nat Qwen3-8B trainer/engine log-probability divergence at iteration 2 requires GPU/runtime investigation, since CPU twin tests cannot establish TokenSpeed GPU parity.

No open issues 🎉

What changed in this PR

Adds dense Qwen3 support alongside Qwen3.5 in AORLTA’s rollout, training, and weight-sync pipeline, deriving model-specific behavior from the tokenizer and checkpoint configuration. Qwen3-8B scoring is demonstrated; training remains experimental because of the documented iteration-2 log-probability gap.

Changes:

• Derives control-token IDs and generation prompts from the model’s tokenizer and chat template.
• Supports family-specific model loading, attention checks, and checkpoint naming.
• Adds CPU end-to-end coverage, Qwen3-8B arithmetic configs, and documented validation limits.

File	Description
aorlta/​VALIDATION.md	Records Qwen3 validation results and limitations.
aorlta/​tools/​merge_base_adapter.py	Accepts either supported decoder prefix during adapter export.
aorlta/​tests/​trainer/​test_kernels.py	Tests when linear-attention checks apply.
aorlta/​tests/​orchestrator/​test_wiring_e2e.py	Allows an injectable engine-model loader.
aorlta/​tests/​orchestrator/​test_wiring_e2e_qwen3.py	Adds two-iteration Qwen3 CPU integration coverage.
aorlta/​src/​aorlta/​trainer/​rank_pool.py	Selects the model loader by checkpoint family.
aorlta/​src/​aorlta/​trainer/​kernels.py	Bypasses linear-attention checks for verified full-attention models.
aorlta/​src/​aorlta/​sync/​writer.py	Rejects mixed decoder prefixes and documents Qwen3 layouts.
aorlta/​src/​aorlta/​sync/​reload.py	Clarifies model-dependent reload counts in documentation.
aorlta/​src/​aorlta/​sync/​merge.py	Extends LoRA target-key matching to Qwen3.
aorlta/​src/​aorlta/​rollout/​templates.py	Derives chat control IDs and thinking prefixes from tokenizers.
aorlta/​src/​aorlta/​rollout/​feedback.py	Uses model-specific repair suffixes for token bounds.
aorlta/​src/​aorlta/​rollout/​fake_engine.py	Honors the request’s stop token.
aorlta/​src/​aorlta/​rollout/​engine_pool.py	Passes model-specific stop tokens to self-tests.
aorlta/​src/​aorlta/​rollout/​chain.py	Uses model stop tokens for generation and reply stripping.
aorlta/​src/​aorlta/​orchestrator/​wiring.py	Wires tokenizer-derived stops and validates trainer model families.
aorlta/​src/​aorlta/​model_family.py	Defines and validates supported checkpoint families.
aorlta/​README.md	Documents Qwen3 support and measured limitations.
aorlta/​examples/​qa/​qa_task.py	Removes parser-specific hard-coded stop tokens.
aorlta/​examples/​hip-kernels/​hip_kernels.py	Removes parser-specific hard-coded stop tokens.
aorlta/​configs/​arith-train-qwen3-8b.yaml	Adds a single-GPU Qwen3-8B arithmetic training config.
aorlta/​configs/​arith-score-qwen3-8b.yaml	Adds a Qwen3-8B held-out arithmetic scoring config.

_{Review effort: Balanced}