# Bounded excerpts captured from retained Docker logs on 2026-07-12.
# Earlier healthy-but-semantically-wrong responses were lost on container
# recreation and are deliberately not represented as raw evidence here.

[vLLM, skip FP8 KV layers 0-39, 131072]
non-default args: model=poolside/Laguna-XS-2.1-NVFP4 revision=07133fb3df1cc3111478e24ee71a823a598c8c2f max_model_len=131072 max_num_seqs=5 kv_cache_dtype=fp8 kv_cache_dtype_skip_layers=0..39
Using FLASH_ATTN attention backend
Model loading took 20.18 GiB memory and 18.326931 seconds
torch.compile took 30.53 s in total
Profiling CUDA graph memory: PIECEWISE=4 (largest=8), FULL=3 (largest=4)
# The retained log ends here; no API-ready event followed.

[SGLang, flashinfer_trtllm MoE runner, 262144]
Load weight end. elapsed=50.18 s, type=LagunaForCausalLM, quant=compressed-tensors, avail mem=73.41 GB, mem usage=20.32 GB.
KV Cache is allocated. dtype: torch.bfloat16, #tokens: 406540
KV Cache is allocated. dtype: torch.bfloat16, #tokens: 508175
Running FlashInfer autotune ... sm120
Scheduler hit an exception
File "compressed_tensors_w4a4_nvfp4_moe.py", line 333, in apply_weights
  assert layer.routing_method_type is not None
AssertionError