Skip to content

[llm] add kv cache to eval #3162

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 1 commit into from
Closed

[llm] add kv cache to eval #3162

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
30 changes: 25 additions & 5 deletions examples/models/llama2/eval_llama_lib.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -41,13 +41,15 @@ def __init__(
model: nn.Module,
tokenizer: Union[SentencePieceTokenizer, Tiktoken],
max_seq_length: Optional[int] = None,
use_kv_cache: bool = False,
):
device = "cuda" if torch.cuda.is_available() else "cpu"
super().__init__(device=device)
self._model = model
self._tokenizer = tokenizer
self._device = torch.device(device)
self._max_seq_length = 2048 if max_seq_length is None else max_seq_length
self._use_kv_cache = use_kv_cache

@property
def eot_token_id(self):
Expand Down Expand Up @@ -83,7 +85,15 @@ def tok_decode(self, tokens):
return decoded

def _model_call(self, inps):
return self._model(inps)
if self._use_kv_cache:
result_logits = []
for pos in range(self._max_seq_length):
pos_tensor = torch.tensor([pos], dtype=torch.int64)
logits = self._model(inps[:, pos : pos + 1], pos_tensor)
result_logits.append(logits)
return torch.cat(result_logits, dim=1)
else:
return self._model(inps)

def _model_generate(self, context, max_length, eos_token_id):
raise Exception("unimplemented")
Expand All @@ -107,13 +117,22 @@ def __init__(
from executorch.extension.pybindings.portable_lib import _load_for_executorch

self._et_model = _load_for_executorch(self._model)
self._use_kv_cache = self._et_model.run_method("use_kv_cache")[0]

def _model_call(self, inps):
# Given inps (tokens), return the logits from a single forward call
# inps: Tensor of shape (1, max_seq_len - 1)
# logits: Tensor of shape (1, max_seq_len - 1, 32000)
result = self._et_model.forward((inps,))
return result[0]
# logits: Tensor of shape (1, max_seq_len - 1, vocab_size)
if self._use_kv_cache:
result_logits = []
for pos in range(self._max_seq_length):
pos_tensor = torch.tensor([pos], dtype=torch.int64)
logits = self._et_model.forward((inps[:, pos : pos + 1], pos_tensor))
result_logits.append(logits[0])
return torch.cat(result_logits, dim=1)
else:
result = self._et_model.forward((inps,))
return result[0]


class ETRunnerEvalWrapper(GPTFastEvalWrapper):
Expand All @@ -139,7 +158,7 @@ def _model_call(self, inps):

# Example:
# inps: Tensor of shape (1, N)
# logits: Tensor of shape (1, N, 32000)
# logits: Tensor of shape (1, N, vocab_size)
pass


Expand Down Expand Up @@ -225,6 +244,7 @@ def gen_eval_wrapper(
model=model,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
use_kv_cache=args.use_kv_cache,
)


Expand Down