Add support for GPT-NeoX (Pythia) #50
Conversation
|
NOTE: Dolly V2 is not supported by this PR, because it uses Bfloat16, which some of our kernel do not support. It will be added by another PR. |
cacheflow/models/gpt_neox.py
Outdated
|
|
||
| def initialize_dummy_weights(self) -> None: | ||
| for param in self.state_dict().values(): | ||
| param.data.uniform_(-0.1, 0.1) |
There was a problem hiding this comment.
Nit: The U(-0.1, 0.1) initialization will lead to many out-of-ranges and NaNs during the model execution. Maybe use a smaller range like U(-1e-5, 1e-5)?
There was a problem hiding this comment.
The (-0.1, 0.1) initialization actually works. However, to be cautious, I changed the range to (-1e-3, 1e-3).
| self.max_position = 8192 | ||
| self.tie_word_embeddings = config.tie_word_embeddings | ||
|
|
||
| def get_param_size(self) -> int: |
There was a problem hiding this comment.
Can we get the parameter size by counting the actual parameters after the models get initialized? Use some code like the following:
mem_params = sum([param.nelement()*param.element_size() for param in model.parameters()])There was a problem hiding this comment.
Good idea. Let's do that in another PR.
| return dtype_size * total | ||
|
|
||
| def get_max_num_gpu_blocks( | ||
| def get_max_act_size( |
There was a problem hiding this comment.
Similarly, can we profile the actual max activation size by running the model once without any KV cache?
SUMMARY: * "remote push" job for multi-gpu runner. * "remote push" job for single gpu runner. * patches for re-initialization of "ray". found other places in `vllm` where they are passing in `ignore_reinit_error=True`, it just looked like they missed a couple of places. * patch "find" command to only find *.py files starting with "test_". TEST PLAN: runs on remote push --------- Co-authored-by: andy-neuma <[email protected]>
* update quark quantizer command * typo * Using scaled_mm for untuned gemm * remove comment * fix yapf
### What this PR does / why we need it? This PR updates the dependency version of vllm-ascend on torch-npu, so that the vllm-ascend can be installed in a later version environment (like to torch-npu 2.6.0rc1), ### Does this PR introduce _any_ user-facing change? No ### How was this patch tested? CI Test Signed-off-by: ji-huazhong <[email protected]>
[Bug] Fix test for Blackwell
Co-authored-by: dengyunyang <[email protected]>
2 participants
This PR adds support for the GPT-NeoX (Pythia) model, which is the backbone of many popular models including Dolly V2, Stable-LM, and Open Assistant.