Fix MistralIntegrationTest
#31231
Fix MistralIntegrationTest
#31231
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. |
| @@ -526,7 +526,7 @@ def test_model_7b_logits(self): | |||
| # Note: Key 9 is currently set for MI300, but may need potential future adjustments for H100s, | |||
| # considering differences in hardware processing and potential deviations in output. | |||
| EXPECTED_SLICE = { | |||
| 7: torch.tensor([-5.8781, -5.8616, -0.1052, -4.7200, -5.8781, -5.8774, -5.8773, -5.8777, -5.8781, -5.8780, -5.8781, -5.8779, -1.0787, 1.7583, -5.8779, -5.8780, -5.8783, -5.8778, -5.8776, -5.8781, -5.8784, -5.8778, -5.8778, -5.8777, -5.8779, -5.8778, -5.8776, -5.8780, -5.8779, -5.8781]), | |||
| 7: torch.tensor([-5.8828, -5.8633, -0.1042, -4.7266, -5.8828, -5.8789, -5.8789, -5.8828, -5.8828, -5.8828, -5.8828, -5.8828, -1.0801, 1.7598, -5.8828, -5.8828, -5.8828, -5.8828, -5.8828, -5.8828, -5.8828, -5.8828, -5.8828, -5.8828, -5.8828, -5.8828, -5.8828, -5.8828, -5.8828, -5.8828]), | |||
There was a problem hiding this comment.
should update in #29905 but forgot
| @slow | ||
| @require_bitsandbytes | ||
| def test_model_7b_generation(self): | ||
| EXPECTED_TEXT_COMPLETION = { | ||
| 7: "My favourite condiment is 100% ketchup. I love it on everything. I'm not a big", | ||
| 7: "My favourite condiment is 100% ketchup. I’m not a fan of mustard, mayo,", |
There was a problem hiding this comment.
should update in #29905 but forgot
| del model | ||
| backend_empty_cache(torch_device) | ||
| gc.collect() | ||
|
|
There was a problem hiding this comment.
not help and worse cause some GPU OOM in subsequent tests
There was a problem hiding this comment.
Happy to have this deleted but very confused why this would cause OOM 😭
There was a problem hiding this comment.
Got to say I am confused too. torch.cuda.empty_cache is not really a magic
empty_cache() doesn’t increase the amount of GPU memory available for PyTorch.
but I was not expecting it would have undesired side-effect like this (even if it is not helpful).
I don't check if del model and gc.collect() plays a role here though.
There was a problem hiding this comment.
Out of curiosity and keep info here for the record:
- it is
test_model_7b_long_promptgets OOM.- previously with those empty cache, at the beginning of
test_model_7b_long_prompt,nvidia-smishows150MiB / 15360MiBwhich looks nice but we get OOM afterward inside this test - wihout empty cache,
nvidia-smishows9066MiB/ 15360MiBwhich looks not great but we DON'T get OOM afterward inside this test
- previously with those empty cache, at the beginning of
It's very mysterious to me.
| @@ -635,7 +622,7 @@ def test_speculative_generation(self): | |||
| # Note: Key 9 is currently set for MI300, but may need potential future adjustments for H100s, | |||
| # considering differences in hardware processing and potential deviations in generated text. | |||
| EXPECTED_TEXT_COMPLETION = { | |||
| 7: "My favourite condiment is 100% Sriracha. I love the heat, the tang and the fact costs", | |||
| 7: "My favourite condiment is 100% ketchup. I love it on everything. I’m not a big", | |||
| del model | ||
| backend_empty_cache(torch_device) | ||
| gc.collect() | ||
|
|
There was a problem hiding this comment.
Happy to have this deleted but very confused why this would cause OOM 😭
* fix * fix * fix * fix --------- Co-authored-by: ydshieh <[email protected]>
What does this PR do?
test_speculative_generation: it was failed due to 9efec11 (Jan 19 2024) then 2e27291 (May 13 2024). I would trust those 2 PRs and simply update the expected outputs (cc @gante )test_model_7b_generationandtest_model_7b_logits: it is from my PR Fix slow tests for important models to be compatible with A10 runners #29905 where I changeddtypeand/orload_in_4bitinfrom_pretrainedin the tests but forgot to update the expected output values for T4.Also,
this is not helping and worse, with it, we actually get
GPU OOM(here fortest_model_7b_long_prompt_sdpa). After removing them, the tests are all passing now.