Merged
Conversation
stamalakhov
force-pushed
the
save_vram
branch
2 times, most recently
from
0130e7e to
2f97b87
Compare
mhs4670go
reviewed
Apr 14, 2026
| unit="batch", | ||
| disable=not gptq_conf.show_progress, | ||
| ): | ||
| device = next(model.parameters()).device |
Contributor
There was a problem hiding this comment.
(minor) Getting device is recomputed inside the inner loop. It could be moved outside the loop to avoid redundant work and improve readability.
Comment on lines
+126
to
+132
| self.cache_kwargs[k].append( | ||
| v.cpu() | ||
| if isinstance(v, torch.Tensor) | ||
| else (v[0].cpu(), v[1].cpu()) | ||
| if isinstance(v, tuple) | ||
| else None | ||
| ) |
Contributor
There was a problem hiding this comment.
We used to preserve kwargs as-is, but this patch now converts only tensors and 2-element tuples, and replaces everything else with None. That can change input semantics.
I think this should be implemented with a recursive “move tensors only” helper that preserves the original container structure.
# let's add this function to tico/utils/utils.py
def move_to_device(obj, device):
"""
Recursively move tensors inside a nested structure to the given device.
Non-tensor objects are preserved as-is.
"""
if isinstance(obj, torch.Tensor):
return obj.to(device)
elif isinstance(obj, tuple):
return tuple(move_to_device(x, device) for x in obj)
elif isinstance(obj, list):
return [move_to_device(x, device) for x in obj]
elif isinstance(obj, dict):
return {k: move_to_device(v, device) for k, v in obj.items()}
# preserve everything else (bool, int, None, custom objects, etc.)
return obj
def move_to_cpu(obj):
return move_to_device(obj, "cpu")Then, we can just call it like this.
# after
self.cache_kwargs[k].append(move_to_cpu(v))
cache_kwargs_batch = gather_single_batch_from_dict(self.cache_kwargs, batch_idx)
cache_kwargs_batch = move_to_device(cache_kwargs_batch, device)
cache_args_batch = gather_single_batch_from_list(self.cache_args, batch_idx)
cache_args_batch = move_to_device(cache_args_batch, device)
Contributor
Author
There was a problem hiding this comment.
Ok. Understood. Thank you!
stamalakhov
force-pushed
the
save_vram
branch
2 times, most recently
from
ce1303e to
1e68dda
Compare
mhs4670go
reviewed
Apr 15, 2026
| return padding | ||
|
|
||
|
|
||
| def move_to_device(obj, device): |
Contributor
There was a problem hiding this comment.
How about moving this to tico/utils/utils.py? This function seems to be used in other places as well.
This PR uses gpu memory in GPTQ algorithm only for inference to reduce gpu memory usage. TICO-DCO-1.0-Signed-off-by: s.malakhov <s.malakhov@partner.samsung.com>
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
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 uses gpu memory in GPTQ algorithm only for inference to reduce gpu memory usage.
It will make it possible to use large number of samples on a gpu with constrained memory.
Sample run on 256 samples for TinyLlama/TinyLlama-1.1B-Chat-v1.0' on 8Gb GPU
./ccex test --include-internal -k quantization.algorithm.test_gptq
TICO-DCO-1.0-Signed-off-by: s.malakhov s.malakhov@partner.samsung.com