Each troubleshooting entry: the symptom, the first check, and where the fix is
Twenty-four symptoms in five groups: the server does not start, requests fail, the server is slow, the output is wrong, and the benchmark is wrong. Each row gives the first log line or metric to check and the section or action that fixes it.
WHAT YOU SEE
CHECK FIRST
THEN
the server does not start
download hangs
cache folder size
hf download
model not supported
architecture in error
supported list
driver or PTX error
error text
update driver
KV cache and max length
length in error
2.2
free memory below share
nvidia-smi
2.2
crash with FP8 KV cache
backend line
2.4
DGX Spark out of memory
free -g
2.3
slow first start, GB10
Model loading took
2.3
compile step fails
torch.compile error
clear cache
multi-GPU start hangs
NCCL_DEBUG=TRACE
6.1
shorter max length
derived length
2.6
requests fail
HTTP 400, context length
counts in error
shorter request
every request fails
EngineCore error
restart
the server is slow
first token takes seconds
waiting, KV usage
3.1
long gaps, many users
P99 against median
3.3
preemptions rise
num_preemptions
3.1
n-gram gives less
async warning
4.2
eager start slow
first requests
4.3
NVFP4 slow on GB10
TPOT against 0.28
pin release
the output is wrong
answer starts with think
content field
3.10
one token repeated
cache hit, spec
turn one off
invalid JSON with spec
run without spec
3.5
output changes per run
same request alone
3.8
the benchmark is wrong
second run faster
prefix cache hits
5.1