[CI] Fixing some AMD failures #42879
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. |
remi-or
left a comment
remi-or
left a comment
There was a problem hiding this comment.
A comment (repeated twice) to address + I will run the slow tests
|
|
||
| else: | ||
| batch_size, seq_length, _ = inputs_embeds.shape | ||
| batch_size, seq_length = input_ids.shape |
There was a problem hiding this comment.
I think input_ids are optional, as is input_embeds -- is there an argument that is not that we could infer those dimensions with? Otherwise I think one of the two inputs_needs to not beNone, so we can get the dims from whichever is notNone`
There was a problem hiding this comment.
Okay, let's do that. We'll try to get the dimensions from input_embeds. If they are None (as was the case for some failing tests), we will then try to infer them from input_ids.
|
|
||
| else: | ||
| batch_size, seq_length, _ = inputs_embeds.shape | ||
| batch_size, seq_length = input_ids.shape |
| out = model(input_ids).logits.float().cpu() | ||
| # Expected mean on dim = -1 | ||
| EXPECTED_MEAN = torch.tensor([[-1.9537, -1.6193, -1.4123, -1.4673, -1.8511, -1.9309, -1.9826, -2.1776]]) | ||
| EXPECTED_MEAN = torch.tensor([[-2.2121, -1.6335, -1.4816, -1.5035, -1.9110, -1.8979, -1.9682, -2.1980]]) |
There was a problem hiding this comment.
I will run the slow test to see if this does not break things on CUDA, otherwise you should add Expectations
There was a problem hiding this comment.
Yes. I remember checking the Nvidia CI, they were the same expectation outputs. But let's see for these exepected means.
|
run-slow: qwen2, qwen2_5_omni, qwen2_5_vl |
|
This comment contains models: ["models/qwen2", "models/qwen2_5_omni", "models/qwen2_5_vl"] |
CI Results✅ No failing test specific to this PR 🎉 ! |
Abdennacer-Badaoui
force-pushed
the
amd-ci-failures
branch
from
ae6c74a to
88c323d
Compare
Abdennacer-Badaoui
force-pushed
the
amd-ci-failures
branch
from
88c323d to
76ebf07
Compare
|
[For maintainers] Suggested jobs to run (before merge) run-slow: qwen2, qwen2_5_omni, qwen2_5_vl |
3 participants
This PR fixes several failures on AMD for Qwen2, Qwen2.5-Omni, and Qwen2.5-VL.