[None][feat] Opt GPT-OSS in to KV cache manager V2 by default

[eopXD]

Description

Opts GPT-OSS in to KVCacheManagerV2 by default.

GPT-OSS applies a sliding window to every other layer (AttentionBlock.__init__, tensorrt_llm/_torch/models/modeling_gpt_oss.py:96), so its KV cache is VSWA with two distinct attention window sizes. V2 groups layers by lifecycle and coalesces buffers within each pool group, sizing the sliding-window and full-attention pools independently instead of statically dividing memory between them.

The change adds GptOssForCausalLM.get_model_defaults() returning kv_cache_config.use_kv_cache_manager_v2=True. This plugs into the per-model auto-selection infrastructure added in #15823:

• use_kv_cache_manager_v2 defaults to the "auto" sentinel (tensorrt_llm/llmapi/llm_args.py:3766).
• ModelLoader merges model defaults, then _resolve_kv_cache_manager_v2_auto (tensorrt_llm/llmapi/llm_utils.py:560) collapses "auto" to a concrete bool.
• An explicit user use_kv_cache_manager_v2 always wins over the model default.
• In disaggregated serving, the existing guard falls back to V1 unless the route is backend=NIXL + transceiver_runtime=PYTHON. GPT-OSS already declares get_preferred_transceiver_runtime() -> "PYTHON", so the NIXL route keeps V2.

GPT-OSS joins hybrid Mamba (NemotronH, Qwen3-Next) and DeepSeek-V4 as models that select V2 by default.

Relationship to #14047

This is being validated alongside #14047 (KVCacheManagerV2 C++ translation), but it has no code dependency on it and is proposed against main on its own. On top of #14047 the flip additionally moves GPT-OSS onto the C++ V2 core, since TLLM_KV_CACHE_MANAGER_V2_BACKEND defaults to cpp there — accuracy/perf runs are being done with both backend settings so a regression can be attributed to the manager switch versus the language switch.

Known limitation: two-model Eagle3

V2 does not split the KV cache budget between the target and draft managers (tensorrt_llm/_torch/pyexecutor/_util.py:1681-1689, assert at :1708). Three GPT-OSS Eagle3 accuracy tests parametrize one_model=[True, False] without setting the flag, so their two_model variants would have silently picked up V2:

• TestGPTOSS::test_eagle3_guided_decoding_4gpus
• TestGPTOSS::test_eagle3_2gpus
• TestGPTOSS::test_eagle3_1gpu

These now pass use_kv_cache_manager_v2="auto" if one_model else False, which pins V1 for two_model (preserving current behavior) while leaving one_model on "auto" so it exercises the new default. This mirrors the existing pytest.skip("KVCacheManagerV2 not compatible with two-model overlap scheduling") in test_eagle3_4gpus and test_eagle3_vswa_reuse_4gpus.

DFlash is unaffected: SpeculativeDecodingMode.has_draft_model() excludes DFLASH, so no separate draft engine is created.

Scope is intentionally limited to the manager selection. Other V2 tuning knobs GPT-OSS could take (tokens_per_block, enable_swa_scratch_reuse — both set by DeepSeek-V4) are deliberately left out pending perf measurement.

Test Coverage

New:

• tests/unittest/_torch/modeling/test_modeling_gpt_oss.py::test_gpt_oss_model_defaults — asserts the defaults dict, mirroring test_modeling_deepseekv4.py::test_deepseek_v4_model_defaults.

Existing coverage that exercises this path:

• "auto"-resolution precedence (explicit user value wins, unknown model falls back to False) is covered generically in tests/unittest/llmapi/test_llm_args.py:451-487.
• TestGPTOSS::test_w4_1gpu, test_w4_4gpus, test_eagle3_4gpus, test_eagle3_vswa_reuse_4gpus already parametrize v2_kv_cache explicitly; they pass the flag, so they are unaffected by the default change and continue to cover both managers.
• The one_model variants of the three tests modified here newly run through the "auto" resolution path.

Validation status: formatting (yapf/isort) and baseline-gated lint pass locally. Draft until the GPT-OSS accuracy (GSM8K / GPQA) and perf runs complete; results will be posted here.

PR Checklist

Please review the following before submitting your PR:

• PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.

• PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.

• Test cases are provided for new code paths (see test instructions)

• If PR introduces API changes, an appropriate PR label is added - either api-compatible or api-breaking. For api-breaking, include BREAKING in the PR title.

• Any new dependencies have been scanned for license and vulnerabilities

• CODEOWNERS updated if ownership changes

• Documentation updated as needed

• Update tava architecture diagram if there is a significant design change in PR.

• The reviewers assigned automatically/manually are appropriate for the PR.

• Please check this after reviewing the above items as appropriate for this PR.

GitHub Bot Help

To see a list of available CI bot commands, please comment /bot help.