fix(coli): --gpu on DeepSeek V4 refused an install with only the DeepGEMM DLL

[kevin9327]

Symptom

On Windows, a DeepSeek V4 install with coli_cuda_dsv4_dg.dll next to deepseek_v4.exe but no coli_cuda_dsv4.dll gets two contradicting answers. That is the DeepGEMM build docs/deepseek-v4.md lists for RTX 50-series:

$ coli doctor --model v4_tiny
[  ok] engine.binary      engine executable is ready
[  ok] accelerator.gpu    GPU engine and devices are available

$ coli run --gpu 0 --model v4_tiny "x"
--gpu needs the CUDA build: build coli_cuda_dsv4.dll (make cuda-dsv4-dll)
(exit 1)

--vram refuses the same way. The engine itself would have loaded that DLL.

Root cause

dsv4_cuda_available in c/coli tests only coli_cuda_dsv4.dll on win32. Two other places accept either DLL:

• The engine's loader, backend_loader_dsv4.c dsv4_cuda_load, tries coli_cuda_dsv4_dg.dll first and coli_cuda_dsv4.dll second. Either one starts the tier.
• doctor.cuda_linkage accepts either name for a [DSV4 CUDA] engine.

Fix

On win32, dsv4_cuda_available accepts either of the two names the loader tries, the same pair doctor accepts. Linux and macOS are unchanged, and so is the build hint.

Tests

In c/tests/test_env_defaults.py (Dsv4CudaDetectTest):

• test_win32_deepgemm_dll_alone_detected: only coli_cuda_dsv4_dg.dll next to the engine is detected.
• test_win32_unrelated_dll_rejected: a lone coli_cuda.dll (the GLM backend) is still not accepted.
• _detect takes the DLL name as a parameter. Existing cases keep the old default.

Fail-before on unfixed dev:

FAIL: test_win32_deepgemm_dll_alone_detected (tests.test_env_defaults.Dsv4CudaDetectTest.test_win32_deepgemm_dll_alone_detected)
AssertionError: False is not true
FAILED (failures=1)

Verification

• tests.test_cli_output, test_cuda_binary_engine, test_env_defaults, test_stop_scope, test_v4_cli (every module that exercises the launcher's GPU gating): 92 tests OK (1 skipped), in the cp949 locale and with PYTHONUTF8=1.
• End to end, Windows, with deepseek_v4.exe from make -C c deepseek-v4 (gcc 16.2), the tiny checkpoint from tools/make_deepseek_v4_tiny.py, and an empty coli_cuda_dsv4_dg.dll next to the engine:
  • Before the fix, coli run --gpu 0 --ngen 4 --model v4_tiny "x" exits 1 with the refusal above.
  • After the fix, the same command reaches the engine. The engine tries both names, finds no usable backend (the file is a placeholder) and continues on the CPU:

    [DSV4 CUDA] no usable backend DLL (coli_cuda_dsv4_dg.dll / coli_cuda_dsv4.dll); GPU tier disabled (CPU path remains active).
    v4_tokens prompt=1 generated=4 total=5 ...
    

    It exits 0.

🤖 Generated with Claude Code