[CI] Drop the redundant and broken tc_piecewise tests

[nvpohanh]

[by Claude Code]

Motivation

The tc_piecewise tests for Llama and Qwen2 models fail on main while capturing the prefill CUDA graph:

torch._dynamo.exc.Unsupported: Unsupported object.__new__ user-defined class construction
  class=<class 'sglang.srt.layers.layer_boundary.residual.stream.Contribution'>,
  error=object.__new__(Contribution) is not safe, use Contribution.__new__()

tc_piecewise compiles the decoder with fullgraph=True. Since the Qwen2 and Llama decoders were built from stage boundaries (#42307, #42308), their forward constructs msgspec.Struct layer-boundary values that Dynamo cannot construct. Failing runs on main:

• Weekly TestPCGWithDFlash (Llama-3.1-8B): https://github.com/sgl-project/sglang/actions/runs/37166616108/job/111330769047
• Weekly TestPCGWithNGRAM (Qwen2.5-Coder-7B): https://github.com/sgl-project/sglang/actions/runs/37166616108/job/111330769003. TestPCGWithSTANDALONE (Llama) did not run in that job, because the run stopped at the NGRAM failure; it fails the same way locally (see below).
• Nightly TestPiecewiseCudaGraphQwen25VL: https://github.com/sgl-project/sglang/actions/runs/37131301136/job/111226858313
• The same file in the per-commit AMD suite stage-b-test-1-gpu-large-amd, which fails with the same error and cancels the sibling partitions: https://github.com/sgl-project/sglang/actions/runs/37253886036/job/111590592756

Per discussion with @ch-wan, tc_piecewise will be removed soon (#41634), so these tests should move to breakable or be removed rather than be fixed.

Modifications

Test	Change	Why
piecewise/test_pcg_with_speculative_decoding_dflash.py (weekly)	Deleted	Under breakable it duplicates the per-commit spec/dflash/test_dflash.py overlap variant: same Llama-3.1-8B + DFlash draft, flashinfer, page size 1, 64 running requests, --mem-fraction-static 0.7, decode batch sizes 1-64, and thresholds 0.75 / 2.8.
TestPCGWithSTANDALONE (weekly)	Deleted	Duplicates the per-commit spec/test_spec_standalone.py: same Llama-3.1-8B + 3.2-1B pair, with stricter thresholds (0.69 / 3.6).
TestPCGWithNGRAM (weekly)	Deleted	Duplicates the per-commit spec/test_spec_ngram.py (+ _extra): same Qwen2.5-Coder-7B, with stricter thresholds (0.79 / 1.8).
TestPCGWithMTP (weekly)	Moved to breakable/test_bcg_with_speculative_decoding_extra.py as TestBCGWithMTP, config unchanged, est_time 450 -> 240	It never selected tc_piecewise and already ran on breakable; it passed in the weekly job above (score 0.97, accept length 3.46). It is the only Qwen3.5-35B-A3B FP8 TP2 MTP check.
piecewise/test_piecewise_cuda_graph_support_1_gpu.py (CUDA nightly + AMD per-commit)	Deleted	Both of its tests exist only to exercise tc_piecewise: GSM8K with tc_piecewise forced, and an embedding comparison between tc_piecewise and no prefill graph. It cannot move to breakable, because Qwen2.5-VL is not on the multimodal breakable allowlist. It fails on both CUDA and AMD.

piecewise/test_pcg_with_speculative_decoding.py (EAGLE3 on Qwen3-30B-A3B) is left for #41634. It still passes because Qwen3-MoE is not built from stage boundaries yet, and its breakable twin breakable/test_bcg_with_speculative_decoding.py already runs per-commit.

Accuracy Tests

This PR changes tests only. Validation:

• pre-commit run --all-files passes.
• ci_register.collect_tests registers MTP as weekly / 4-gpu-h100 / 240 s. No registered test references the removed files or classes.
• On 1x H100 80GB HBM3 (lmsysorg/sglang:dev-cu13, SGLang f70e8c6, torch 2.14.1+cu130), the deleted tc_piecewise tests all fail with the error above. The same configs on breakable (the prefill backend their per-commit equivalents use) pass:

Config	tc_piecewise	breakable: GSM8K score / avg accept length
DFlash (Llama-3.1-8B)	❌ Contribution	✅ 0.845 / 3.92
STANDALONE (Llama-3.1-8B + 3.2-1B)*	❌ Contribution	✅ 0.870 / 3.33
NGRAM (Qwen2.5-Coder-7B)	❌ Contribution	✅ 0.780 / 1.93

* The draft was the ungated mirror unsloth/Llama-3.2-1B-Instruct, because the local HF token cannot access meta-llama/Llama-3.2-1B-Instruct.

Speed Tests and Profiling

N/A (test-only change).

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

🤖 Generated with Claude Code

---

CI States

Latest PR Test (Base): ❌ Run #37301553658
Latest PR Test (Extra): ❌ Run #37301553150
Latest PR Test (AMD ROCm 10): ❌ Run #37301553443

[nvpohanh]

/rerun-test test/registered/cuda_graph/breakable/test_bcg_with_speculative_decoding_extra.py test/registered/spec/dflash/test_dflash.py test/registered/spec/test_spec_standalone.py test/registered/spec/test_spec_ngram.py test/registered/spec/test_spec_ngram_extra.py test/registered/spec/test_spec_standalone_extra.py test/registered/e2e/speculative/test_dflash_domino.py

[github-actions]

🚀 4-gpu-h100 (1 test): ✅ View workflow run

cd test/ && python3 registered/cuda_graph/breakable/test_bcg_with_speculative_decoding_extra.py

🚀 1-gpu-5090 (1 test): ✅ View workflow run

cd test/ && python3 registered/spec/dflash/test_dflash.py

🚀 1-gpu-h100 (4 tests): ✅ View workflow run

cd test/ && python3 registered/spec/test_spec_standalone.py
cd test/ && python3 registered/spec/test_spec_ngram.py
cd test/ && python3 registered/spec/test_spec_ngram_extra.py
cd test/ && python3 registered/spec/test_spec_standalone_extra.py

🚀 2-gpu-h100 (1 test): ✅ View workflow run

cd test/ && python3 registered/e2e/speculative/test_dflash_domino.py