Fix llama conversion with smooth quant #1650
Closed
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
without this it errors out with:
Traceback (most recent call last):
File "/app/tensorrt_llm/examples/llama/convert_checkpoint.py", line 456, in <module>
main()
File "/app/tensorrt_llm/examples/llama/convert_checkpoint.py", line 448, in main
convert_and_save_hf(args)
File "/app/tensorrt_llm/examples/llama/convert_checkpoint.py", line 353, in convert_and_save_hf
LLaMAForCausalLM.quantize(args.model_dir,
File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/models/llama/model.py", line 405, in quantize
convert.quantize(
File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/models/llama/convert.py", line 1395, in quantize
weights = load_weights_from_hf(
File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/models/llama/convert.py", line 1437, in load_weights_from_hf
weights = convert_hf_llama(
File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/models/llama/convert.py", line 1089, in convert_hf_llama
convert_layer(l)
File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/models/llama/convert.py", line 725, in convert_layer
get_tllm_linear_sq_weight(int8_weights,
File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/models/llama/convert.py", line 610, in get_tllm_linear_sq_weight
results[prefix + 'per_channel_scale'] = torch.Tensor([
ValueError: only one element tensors can be converted to Python scalars
we can also check the shapes:
cur_per_channel_value.shape -> torch.Size([6144])
col_shape -> [1, 6144]
so it's clear that we meant to convert the tensor without []
with these the model works and provides sensible output
Merged
Member
Contributor
Author
|
hi @kaiyux great, thank you! I think only the first commit was integrated, but the two others were not, but they are also required -- although they fix the error which would happen when running the engine. I'm experimenting with smooth quant llama 3 right now and need all commits to get it working. Do you mind having another look? |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR fixes a few errors which appear when following the README https://github.com/NVIDIA/TensorRT-LLM/blob/main/examples/llama/README.md#smoothquant on current latest commit.
Note: the first commit looks quite obvious (although I'm not sure how this could have worked before), while I'm less sure about the second, I just just going by error messages during engine conversion, there might be a better place for the fix. So feel free to treat this as a bug report instead. I verified that engine built in this way has reasonable outputs and expected performance. The model I was testing this on is Mistral 7B (mistral-7b-v0.1-instruct) but I assume other llama 2 and 3 should also work (didn't get to llama 3 yet).