huggingface/text-generation-inference
5
0
Fork 0
mirror of https://github.com/huggingface/text-generation-inference.git synced 2025-04-24 00:12:08 +00:00
Code Issues Packages Projects Releases Wiki Activity

Clear previous hpu_graphs when graph shape changed to save memory (#176)

Browse Source
This commit is contained in:
BaihuiJin 2024-07-11 21:19:17 +08:00 committed by GitHub
parent fff1d4f86f
commit aac547dd82
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 4 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
server/text_generation_server/models/causal_lm.py
View File

@ -602,6 +602,7 @@ class CausalLM(Model):
dtype: Optional[torch.dtype] = None, dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False, trust_remote_code: bool = False,
): ):
self.prev_bs = 0
if use_medusa: if use_medusa:
raise RuntimeError("Medusa decoding is not enabled for AutoModel") raise RuntimeError("Medusa decoding is not enabled for AutoModel")
@ -965,6 +966,9 @@ class CausalLM(Model):
batch = batch.__class__.recombine([batch], self.tokenizer.pad_token_id) batch = batch.__class__.recombine([batch], self.tokenizer.pad_token_id)
scenario = 'PREFILL' if prefill else 'GENERATE' scenario = 'PREFILL' if prefill else 'GENERATE'
if self.enable_hpu_graph and batch.batch_size != self.prev_bs:
self.model.clear_cache()
self.prev_bs = batch.batch_size
dbg_trace( dbg_trace(
scenario, f'bs:{batch.batch_size} num_reqs:{len(batch.requests)} seq_len:{batch.seq_length} padding:{batch.right_padding}') scenario, f'bs:{batch.batch_size} num_reqs:{len(batch.requests)} seq_len:{batch.seq_length} padding:{batch.right_padding}')
assert batch.right_padding > 0, 'No more room for next token!' assert batch.right_padding > 0, 'No more room for next token!'