fix(autotune): coli tune swept GLM-only CUDA and NUMA knobs on every engine

[kevin9327]

Symptom

coli tune offers four knobs to every engine that only GLM-5.2 reads:

• cuda-pipe-1, cuda-pipe-2 and cuda-sync, whenever the plan lists a GPU;
• numa-on, on a multi-socket host.

On any other engine each of those candidates runs the baseline again. It costs a full engine start and a full request rotation, and timing noise can be accepted as a gain and written into the saved profile.

coli tune on the tiny Qwen3.6 fixture (Windows, RTX 5080 visible to the planner, CPU-only qwen36.exe), unfixed dev:

  · baseline cap=8 (1/1, prompt 1/2)
  · omp-8 cap=8 (1/1, prompt 1/2)
  · cuda-pipe-2 cap=8 (1/1, prompt 1/2)
  · cuda-sync cap=8 (1/1, prompt 1/2)
  · confirm-winner cap=8 (1/1, prompt 1/2)
  · confirm-baseline cap=8 (1/1, prompt 1/2)

  ✓ accepted cuda-pipe-2: 491.44 → 838.89 tok/s (+70.7%)
  env: OMP_NUM_THREADS=8 COLI_CUDA_PIPE=2

qwen36 never reads COLI_CUDA_PIPE. The reported +70.7% is omp-8 plus noise, and the profile now carries a knob that does nothing.

Root cause

candidate_steps checks arch for PIPE and DIRECT (#898), but not for these four:

knob	where it is read
COLI_CUDA_PIPE	colibri.c only (g_cuda_pipe, line 11098)
COLI_CUDA_ASYNC	expert_group_impl in backend_cuda.cu. That function is reached through coli_cuda_expert_group / _pinned, and colibri.c is their only caller. qwen36 and qwen38 use the _issue/_take pair, which does not read it. olmoe, glm53, deepseek_v41 and inkling do not link backend_cuda.o.
COLI_NUMA	colibri.c only (numa_init, which mbinds colibri's own expert slabs)

The docstring said COLI_NUMA was read by every engine. git grep finds no other reader.

Fix

• Sweep numa-on and the CUDA candidates only when arch == "glm", the same gate PIPE and DIRECT already have.
• Correct the docstring. OMP_NUM_THREADS stays universal.
• GLM's candidate list is unchanged.

Two commits: the CUDA pair, then COLI_NUMA.

Tests

In c/tests/test_autotune.py:

• test_cuda_and_numa_knobs_are_not_swept_on_other_engines uses a plan with a GPU and two sockets.
  • GLM still gets numa-on, cuda-pipe-1, cuda-pipe-2 and cuda-sync.
  • glm53, deepseek_v4, deepseek_v41, kimi, inkling, olmoe, qwen36 and qwen38 get no numa-* or cuda-* candidate and no COLI_NUMA or COLI_CUDA_* key, and keep their omp-* candidate.
  • The existing [Bug]: coli tune banners DeepSeek V4 but calibrates with the GLM engine and fails #898 test uses one socket and no GPU, so it could not see this.
• test_sibling_tune_with_a_gpu_and_two_sockets_launches_only_omp runs run_tune for qwen36 through the existing FakeServeEngine, with the same plan. The candidates are baseline and omp-4, no launch sees any of the three variables, and there are 4 engine launches.

Fail-before on unfixed dev:

AssertionError: True is not false : {'omp-4': {'OMP_NUM_THREADS': '4'}, 'numa-on': {'COLI_NUMA': '1'}, 'cuda-pipe-1': {'COLI_CUDA_PIPE': '1'}, 'cuda-pipe-2': {'COLI_CUDA_PIPE': '2'}, 'cuda-sync': {'COLI_CUDA_ASYNC': '0'}}
...
AssertionError: Lists differ: ['baseline', 'omp-4', 'numa-on', 'cuda-pipe-1', 'cuda-pipe-2', 'cuda-sync'] != ['baseline', 'omp-4']
FAILED (failures=9)

That is 8 subtests plus the run_tune test.

Verification

• tests.test_autotune, tests.test_autotune_engine_output_locale, tests.test_v4_cli (every test module that imports autotune): 47 tests OK, in the cp949 locale and with PYTHONUTF8=1.
• End to end, Windows, real engines on tiny fixtures, coli tune --tokens 4 --repeats 1 (fixtures from tools/make_qwen36_tiny.py + convert_qwen36.py, tools/make_olmoe_tiny.py + convert_olmoe_merged.py, tools/make_glm_oracle.py):
  • qwen36_tiny_c: baseline, omp-8, then the gate; no CUDA candidate.
  • olmoe_tiny_c: baseline, omp-8, confirm-winner, confirm-baseline.
  • glm_tiny: baseline, omp-8, cuda-pipe-2, cuda-sync, confirm-winner, confirm-baseline, as before.
  • The host has one socket, so numa-on is covered by the unit tests only.
• Full Python suite (Windows, gcc 16.2, PYTHONUTF8=1, first commit): 1085 tests, 1 failure and 6 errors, none from this change.
  • The 6 errors are test_cuda_test_makefile, which needs make on PATH.
  • The failure is test_kimi_usage_cli.test_no_arguments_points_at_the_launcher_and_fails. On Windows the engine prints coli.cmd chat and the test looks for coli chat; it has nothing to do with autotune.

🤖 Generated with Claude Code