Fix accelerate failing tests #30836
Fix accelerate failing tests #30836
Conversation
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
| @@ -244,6 +244,7 @@ def slow_forward(self, input_states, cache_params: Optional[MambaCache]=None): | |||
| # 2. Convolution sequence transformation | |||
| if cache_params is not None: | |||
| ssm_state = cache_params.ssm_states[self.layer_idx].clone() | |||
There was a problem hiding this comment.
if the cache param is always on the wrong device we are always gonna have to do the transfer here no?
| @@ -1388,7 +1388,7 @@ def forward( | |||
| inputs_embeds, past_key_values_length=past_key_values_length, position_ids=position_ids | |||
| ) | |||
|
|
|||
| hidden_states = inputs_embeds + positions | |||
| hidden_states = inputs_embeds + positions.to(inputs_embeds.device) | |||
There was a problem hiding this comment.
what was the issue with this? It' s a bit weird here, device placement of the inputs is not automatic? Or is the embedding layer alone place on another device than the embed_position?
There was a problem hiding this comment.
inputs_embeds and postions can be on different devices (could be fixed if we find a way to make sure that they are on the same device, however, we would probably need to add another attribute in PretrainedConfig). We can't really move automatically these inputs to the right device since we do the ops in the forward itself. Wrapping them in a nn.Module to perform the op would work but that's not a good solution.
| @@ -962,7 +962,7 @@ class BertModel(BertPreTrainedModel): | |||
| `add_cross_attention` set to `True`; an `encoder_hidden_states` is then expected as an input to the forward pass. | |||
| """ | |||
|
|
|||
| _no_split_modules = ["BertEmbeddings"] | |||
| _no_split_modules = ["BertEmbeddings", "BertLayer"] | |||
There was a problem hiding this comment.
Is this a general solution that will work for everyone? (I mean why was this failing the test)
There was a problem hiding this comment.
Yes that will work for everyone ! I don't know why the layer were not added to _no_split_modules, this is what we do for all models.
|
Just wanna to know if those are failing for quite some time or it's a recent changes (in either |
I'm not sure how long these tests have been failing. However, most of the tests were failing for a good reason (didn't put the right |
* Fix accelerate tests * fix clip * skip dbrx tests * fix GPTSan * fix M2M100Model * same fix as jamba * fix mt5 * Fix T5Model * Fix umt5 model * fix switch_transformers * fix whisper * fix gptsan again * fix siglip recent test * skip siglip tests * wrong place fixed
What does this PR do ?
This PR fixes some tests (~ 30) are that contained in the accelerate_tests flag. Recently, I ran the following workflow to have an overview of the failing tests : https://github.com/huggingface/transformers/actions/runs/9081904684 or the following CLI:
RUN_SLOW=True pytest -m accelerate_tests tests/Related PR #30808
Model fixed:
_preload_module_classesto be used indispatch_model), not sure if it is worth it if this is the only model for now.Tests skipped:
See the run here with most of the tests passing . There are only 3 tests failing but I skipped them in the latest commit (siglipvisionmodel)
cc @ydshieh