Support SM120
Dear Nvidia, stop being assholes and provide better support for SM120 cards. These cards cost a great deal of money and are used by professionals in the field. The lack of support around nvfp4 in this space is a joke. Do better. I know datacenter cards let Jensen buy more leather jackets, but I think he's good for the moment.
vllm already supported.
trt-llm is trash running qwen3.5 even on B200, according to github
care to actually back that up with your config? vLLM does not support it as far as I can tell
care to actually back that up with your config? vLLM does not support it as far as I can tell
because bad quant config
add .experts.mlp.gate and mtp.fc* to ignore list
care to actually back that up with your config? vLLM does not support it as far as I can tell
because bad quant config
add .experts.mlp.gate and mtp.fc* to ignore list
Can you share your config because MTP is not compatible with NFP4 unless I am missing something?
care to actually back that up with your config? vLLM does not support it as far as I can tell
because bad quant config
add .experts.mlp.gate and mtp.fc* to ignore listCan you share your config because MTP is not compatible with NFP4 unless I am missing something?
sorry but I have problems uploading files.
you can look at my comment in this https://huggingface.co/vincentzed-hf/Qwen3.5-397B-A17B-NVFP4/discussions/1
It's astonishing to me that a company with a $4T+ market cap, virtually unlimited cash and 40,000 employees so badly botched:
-Day ONE complete software support for Blackwell
-TRT-LLM support
-Introduction of NVFP4, so breathlessly hyped to by optimized for Blackwell but so far has been a complete turd in actual use.
These are signs of a company getting arrogant for want of competition. I hope they see some soon.