Closed kaiyux closed 3 months ago
Hi, thanks for your hard work, btw. I have spotted huge removal in examples/run.py
: https://github.com/NVIDIA/TensorRT-LLM/commit/db4edea1e1359bcfcac7bbb87c1b639b5611c721#diff-299cb0140ad8f9d286c86ecc32b793b048531e27570675b94e54b57b66b3d7d5. Is it intented?
Sorry for false alarm, these arguments was moved to utils. I didn't spotted it
examples/phi/README.md
max_batch_size
intrtllm-build
command is 256 by default now.max_num_tokens
intrtllm-build
command is 8192 by default now.api
ingptManagerBenchmark
command isexecutor
by default now.bias
argument to theLayerNorm
module, and supports non-bias layer normalization.LLM.generate()
API.SamplingConfig
SamplingParams
with some sampling parameters, seetensorrt_llm/hlapi/utils.py
SamplingParams
instead ofSamplingConfig
inLLM.generate()
API, seeexamples/high-level-api/README.md
GptManager
APImaxBeamWidth
intoTrtGptModelOptionalParams
schedulerConfig
intoTrtGptModelOptionalParams
convert_hf_mpt_legacy
call failure when the function is called in other than global scope, thanks to the contribution from @bloodeagle40234 in #1534.use_fp8_context_fmha
broken outputs (#1539).--ipc=host
notes to installation guide to prevent bus error, seedocs/source/installation/build-from-source-linux.md
anddocs/source/installation/linux.md
(#1538)