[faresobeid]

Removes GPT-OSS (gpt_oss) support. No back-compat shims: a GPT-OSS checkpoint now fails at trainer setup like any other unsupported architecture.

Removed

• trainer/models/gpt_oss/ (model, its own FA4 attention with learned sinks, and its own ring/Ulysses CP paths), tests/unit/train/models/test_gpt_oss.py.
• inference/vllm/gpt_oss_weight_loading.py and its call in inference/patches.py.
• Registry entries, README / docs/advanced.md rows and the GPT-OSS FA4/sinks note.
• openai/gpt-oss tool-parser auto-detection, plus its parser test rows.

Shared code that only existed for GPT-OSS

• layers/attn.py, layers/ulysses_attn.py: drop the GPT-OSS CP substitution hooks. Also drop the learnable_sink and softmax_scale args of ulysses_flash_attn_varlen_func, because GPT-OSS was their only caller. DeepSeek-V4's sinks live in its own kernels and are untouched.
• layers/moe.py: remove expert projection biases (GroupedExperts(bias=...)), the router gate_bias, and the topk_softmax score function.
• layers/expert_compute.py: remove broadcast_expert_bias and the bias adds in the grouped-GEMM forward.
• layers/activations.py: remove clamped_swiglu (DeepSeek-V4 keeps its own ClampedSwiglu with a configurable limit).
• layers/sonic_moe.py: remove the bias-free check, since experts have no biases now.
• trainer/lora.py + layers/lora/multi_moe.py: remove the model_type == "gpt_oss" branch and MultiLoRAGptOssGroupedExperts.
• test_moe.py: drop the clamped-swiglu/bias cases and the bias terms in the expert reference.

Not touched: DeepseekV4Router.forward still has a gate.bias branch copied from the base router. It is dead now. #3854 deletes that copy.

Checks

• ruff check . and ruff format --check .: clean.
• pytest -m "not gpu" tests/unit/orchestrator tests/unit/test_configs.py tests/unit/test_parsers.py tests/unit/train tests/unit/inference: all pass except test_cpu_adam.py::test_native_cpu_adamw_matches_fused_torch_and_preserves_gradients, which fails the same way on main on this CPU box.
• CPU forward + backward check for qwen3_moe, glm4_moe and laguna (tiny configs, fixed seeds, packed seq_lens, SDPA stand-in for the flash kernel). Logits, all parameter grads and state-dict keys are bitwise identical to main. DeepSeek-V4 needs GPU kernels, so it was not run.
• grep across src/tests/configs/examples/docs/skills/scripts/.github: no references left (except the renderer list in docs/algorithms.md, which describes the external renderers package).
• A checkpoint of this family now fails at trainer setup with the usual "has no PrimeRL model implementation" error. No shim.

Conflicts with open PRs

• Chore/remove transformers modeling #3833 (samsja, remove transformers modeling): edits gpt_oss/*, test_gpt_oss.py, lora.py, attn.py, models/__init__.py. Whichever lands second drops the gpt_oss hunks.
• refactor(lora): single-adapter LoRA layers, drop modules_to_save #3855 (LoRA single-adapter): renames MultiLoRAGptOssGroupedExperts → LoRAGptOssGroupedExperts. If this lands first, drop that class and branch there.
• refactor(models): one shared init_buffers_post_meta (stacked on #3833) #3857 (shared init_buffers_post_meta): edits gpt_oss/modeling_gpt_oss.py. Drop the hunk.
• chore(models): sqrtsoftplus in the shared router, drop the DeepseekV4Router copy #3854, feat(trainer): Composer 2-style plausibility filter for router replay #3796, refactor(trainer): add iter_moe_blocks for the MoE layer lookup #3843, chore(models): share Qwen3.5/Qwen3.8 GatedDeltaNet, norm and shared expert in layers/ #3860: touch layers/moe.py / test_moe.py near the removed bias/topk_softmax lines. Small textual conflicts.
• chore(trainer): AFMoE attention subclasses the shared FlashAttention #3859: touches the same CP substitution block in attn.py / ulysses_attn.py.
• feat(inference): router replay on the V2 runner everywhere #3830–feat(inference): P/D per-role KV offload #3832: inference/patches.py, only the two removed GPT-OSS lines at the top of apply_shared_vllm_patches.