[Bug]: RTX 3080 sm_86, WSL2 — "hybrid batched block failed in MoE" on every prompt

Commit or version

fd93c41

Environment

GPU: NVIDIA RTX 3080 10GB, sm_86 (Ampere) OS: Windows 11, WSL2 Ubuntu Colibri: v1.10.2 (fd93c41) Model: DeepSeek-V4-Flash (160GB official checkpoint) CUDA: 13.3 (Windows) / 12.8 (WSL2)

Reproduction steps

1. Build deepseek-v4 engine with CUDA=1 via WSL2
2. Run: CUDA_DENSE=1 python3 coli chat --model /path/to/deepseek-v4-flash --ram 32
3. Type any prompt

Expected behavior

Model generates a response

Actual behavior and logs

request failed: hybrid batched block failed in MoE

Startup log shows GPU loading correctly:
v4_gpu tier=dense-matvec device=0
NVIDIA GeForce RTX 3080 10.7 GB sm_86

But every inference attempt fails immediately with hybrid batched block error regardless of flags tried: CUDA_DENSE=1, COLI_CUDA_MOE_BATCH=0, COLI_CUDA_ATTN_BATCH=0, DSV4_HYBRID=0

RTX 3080 sm_86, WSL2 — "hybrid batched block failed in MoE" on every prompt

GPU: NVIDIA RTX 3080 10GB, sm_86 (Ampere)
OS: Windows 11, running via WSL2 Ubuntu
Colibri: v1.10.2 (fd93c41)
Model: DeepSeek-V4-Flash (160GB official checkpoint)
CUDA: 13.3 (Windows) / 12.8 (WSL2)

GPU loads successfully — startup log shows v4_gpu tier=dense-matvec device=0. But every inference attempt fails with:

request failed: hybrid batched block failed in MoE

Tried the following — all fail with same error:

CUDA_DENSE=1
COLI_CUDA_MOE_BATCH=0
COLI_CUDA_ATTN_BATCH=0
DSV4_HYBRID=0
All combinations of the above

Any workaround for Ampere 10GB?