Add more statistics for spec decoding (#13317)
This commit is contained in:
@@ -1779,6 +1779,9 @@ class TokenizerManager(TokenizerCommunicatorMixin):
|
||||
meta_info["spec_accept_length"] = (
|
||||
recv_obj.completion_tokens[i] / recv_obj.spec_verify_ct[i]
|
||||
)
|
||||
meta_info["spec_accept_token_num"] = accepted_tokens
|
||||
meta_info["spec_draft_token_num"] = total_draft_tokens
|
||||
meta_info["spec_verify_ct"] = recv_obj.spec_verify_ct[i]
|
||||
|
||||
def _calculate_timing_metrics(
|
||||
self,
|
||||
|
||||
Reference in New Issue
Block a user