Inline extend_range accessors and remove the extend_input_len/fill_len properties (#27611)
This commit is contained in:
@@ -381,9 +381,8 @@ def prepare_inputs_for_correctness_test(bench_args, tokenizer, custom_prompts):
|
||||
sampling_params=sampling_params,
|
||||
)
|
||||
req.full_untruncated_fill_ids = req.origin_input_ids
|
||||
req.fill_len = len(req.full_untruncated_fill_ids)
|
||||
req.logprob_start_len = -1
|
||||
req.set_extend_input_len(req.fill_len - len(req.prefix_indices))
|
||||
req.set_extend_range(len(req.prefix_indices), len(req.origin_input_ids))
|
||||
reqs.append(req)
|
||||
|
||||
return input_ids, reqs
|
||||
@@ -395,14 +394,15 @@ def prepare_extend_inputs_for_correctness_test(
|
||||
for i in range(len(reqs)):
|
||||
req: Req = reqs[i]
|
||||
req.full_untruncated_fill_ids.extend(input_ids[i][bench_args.cut_len :])
|
||||
req.fill_len = len(req.full_untruncated_fill_ids)
|
||||
if model_runner is not None:
|
||||
# Use req.req_pool_idx instead of i to handle slot 0 padding correctly
|
||||
req.prefix_indices = model_runner.req_to_token_pool.req_to_token[
|
||||
req.req_pool_idx, : bench_args.cut_len
|
||||
].to(req.prefix_indices.dtype)
|
||||
req.logprob_start_len = -1
|
||||
req.set_extend_input_len(req.fill_len - len(req.prefix_indices))
|
||||
req.set_extend_range(
|
||||
len(req.prefix_indices), len(req.full_untruncated_fill_ids)
|
||||
)
|
||||
return reqs
|
||||
|
||||
|
||||
@@ -428,9 +428,8 @@ def prepare_synthetic_inputs_for_latency_test(
|
||||
sampling_params=sampling_params,
|
||||
)
|
||||
req.full_untruncated_fill_ids = req.origin_input_ids
|
||||
req.fill_len = len(req.full_untruncated_fill_ids)
|
||||
req.logprob_start_len = -1
|
||||
req.set_extend_input_len(req.fill_len - len(req.prefix_indices))
|
||||
req.set_extend_range(len(req.prefix_indices), len(req.origin_input_ids))
|
||||
reqs.append(req)
|
||||
|
||||
return reqs
|
||||
|
||||
Reference in New Issue
Block a user