[mypy] Enable type checking for test directory (#5017)
This commit is contained in:
@@ -393,7 +393,7 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
|
||||
param_data.copy_(loaded_weight)
|
||||
return
|
||||
current_shard_offset = 0
|
||||
shard_offsets = []
|
||||
shard_offsets: List[Tuple[int, int, int]] = []
|
||||
for i, output_size in enumerate(self.output_sizes):
|
||||
shard_offsets.append((i, current_shard_offset, output_size))
|
||||
current_shard_offset += output_size
|
||||
|
||||
@@ -25,24 +25,25 @@ GPTQ_MARLIN_SUPPORTED_SYM = [True]
|
||||
|
||||
|
||||
# Permutations for Marlin scale shuffling
|
||||
def get_scale_perms(num_bits):
|
||||
scale_perm = []
|
||||
def get_scale_perms(num_bits: int):
|
||||
scale_perm: List[int] = []
|
||||
for i in range(8):
|
||||
scale_perm.extend([i + 8 * j for j in range(8)])
|
||||
scale_perm_single = []
|
||||
scale_perm_single: List[int] = []
|
||||
for i in range(4):
|
||||
scale_perm_single.extend(
|
||||
[2 * i + j for j in [0, 1, 8, 9, 16, 17, 24, 25]])
|
||||
return scale_perm, scale_perm_single
|
||||
|
||||
|
||||
def get_pack_factor(num_bits):
|
||||
def get_pack_factor(num_bits: int):
|
||||
assert (num_bits in GPTQ_MARLIN_SUPPORTED_NUM_BITS
|
||||
), f"Unsupported num_bits = {num_bits}"
|
||||
return 32 // num_bits
|
||||
|
||||
|
||||
def marlin_permute_scales(s, size_k, size_n, group_size, num_bits):
|
||||
def marlin_permute_scales(s: torch.Tensor, size_k: int, size_n: int,
|
||||
group_size: int, num_bits: int):
|
||||
scale_perm, scale_perm_single = get_scale_perms(num_bits)
|
||||
if group_size < size_k and group_size != -1:
|
||||
s = s.reshape((-1, len(scale_perm)))[:, scale_perm]
|
||||
|
||||
@@ -1,4 +1,6 @@
|
||||
"""This file is used for /tests and /benchmarks"""
|
||||
from typing import Dict, List
|
||||
|
||||
import numpy
|
||||
import torch
|
||||
|
||||
@@ -11,10 +13,10 @@ import torch
|
||||
#
|
||||
# As a result of this reordering, the vector loads inside the kernel will get the data as it is needed for tensor-core # noqa: E501
|
||||
# (without the need to use ldmatrix instructions) # noqa: E501
|
||||
def get_perms_24(num_bits):
|
||||
perm_list = []
|
||||
def get_perms_24(num_bits: int):
|
||||
perm_list: List[int] = []
|
||||
for i in range(32):
|
||||
perm1 = []
|
||||
perm1: List[int] = []
|
||||
col = i // 4
|
||||
col_o = col // 2
|
||||
for block in [0, 1]:
|
||||
@@ -39,18 +41,18 @@ def get_perms_24(num_bits):
|
||||
|
||||
perm = perm.reshape((-1, len(interleave)))[:, interleave].ravel()
|
||||
perm = torch.from_numpy(perm)
|
||||
scale_perm = []
|
||||
scale_perm: List[int] = []
|
||||
for i in range(8):
|
||||
scale_perm.extend([i * 8 + j for j in [0, 4, 1, 5, 2, 6, 3, 7]])
|
||||
scale_perm_single = []
|
||||
scale_perm_single: List[int] = []
|
||||
for i in range(8):
|
||||
scale_perm_single.extend([8 * i + j for j in [0, 1, 2, 3, 4, 5, 6, 7]])
|
||||
return perm, scale_perm, scale_perm_single
|
||||
|
||||
|
||||
marlin_24_perm = {}
|
||||
marlin_24_scale_perm = {}
|
||||
marlin_24_scale_perm_single = {}
|
||||
marlin_24_perm: Dict[int, torch.Tensor] = {}
|
||||
marlin_24_scale_perm: Dict[int, List[int]] = {}
|
||||
marlin_24_scale_perm_single: Dict[int, List[int]] = {}
|
||||
for num_bits in [4, 8]:
|
||||
perm_24, scale_perm_24, scale_perm_single_24 = get_perms_24(num_bits)
|
||||
marlin_24_perm[num_bits] = perm_24
|
||||
|
||||
@@ -1,4 +1,6 @@
|
||||
"""This file is used for /tests and /benchmarks"""
|
||||
from typing import Dict, List
|
||||
|
||||
import numpy
|
||||
import torch
|
||||
|
||||
@@ -11,10 +13,10 @@ import torch
|
||||
#
|
||||
# As a result of this reordering, the vector loads inside the kernel will get the data as it is needed for tensor-core # noqa: E501
|
||||
# (without the need to use ldmatrix instructions) # noqa: E501
|
||||
def get_perms(num_bits):
|
||||
perm_list = []
|
||||
def get_perms(num_bits: int):
|
||||
perm_list: List[int] = []
|
||||
for i in range(32):
|
||||
perm1 = []
|
||||
perm1: List[int] = []
|
||||
col = i // 4
|
||||
for block in [0, 1]:
|
||||
for row in [
|
||||
@@ -38,19 +40,19 @@ def get_perms(num_bits):
|
||||
|
||||
perm = perm.reshape((-1, len(interleave)))[:, interleave].ravel()
|
||||
perm = torch.from_numpy(perm)
|
||||
scale_perm = []
|
||||
scale_perm: List[int] = []
|
||||
for i in range(8):
|
||||
scale_perm.extend([i + 8 * j for j in range(8)])
|
||||
scale_perm_single = []
|
||||
scale_perm_single: List[int] = []
|
||||
for i in range(4):
|
||||
scale_perm_single.extend(
|
||||
[2 * i + j for j in [0, 1, 8, 9, 16, 17, 24, 25]])
|
||||
return perm, scale_perm, scale_perm_single
|
||||
|
||||
|
||||
marlin_perm = {}
|
||||
marlin_scale_perm = {}
|
||||
marlin_scale_perm_single = {}
|
||||
marlin_perm: Dict[int, torch.Tensor] = {}
|
||||
marlin_scale_perm: Dict[int, List[int]] = {}
|
||||
marlin_scale_perm_single: Dict[int, List[int]] = {}
|
||||
for num_bits in [4, 8]:
|
||||
perm, scale_perm, scale_perm_single = get_perms(num_bits)
|
||||
marlin_perm[num_bits] = perm
|
||||
|
||||
@@ -174,7 +174,7 @@ def _apply_min_tokens_penalty(
|
||||
min_tokens = sampling_params.min_tokens
|
||||
token_ids_to_penalize = sampling_params.all_stop_token_ids
|
||||
if min_tokens > 0 and token_ids_to_penalize:
|
||||
seqs_to_penalize = []
|
||||
seqs_to_penalize: List[int] = []
|
||||
for j, seq_id in enumerate(seq_ids):
|
||||
seq_data = seq_group.seq_data[seq_id]
|
||||
if len(seq_data.output_token_ids) < min_tokens:
|
||||
@@ -285,7 +285,7 @@ def _greedy_sample(
|
||||
same as the length of selected_seq_groups. If the corresponding
|
||||
seq_group has do_sample=False, tuple contains ([], [])
|
||||
"""
|
||||
samples = samples.tolist()
|
||||
samples_lst = samples.tolist()
|
||||
sample_idx = 0
|
||||
results: SampleResultType = []
|
||||
for seq_group in selected_seq_groups:
|
||||
@@ -298,7 +298,7 @@ def _greedy_sample(
|
||||
assert num_parent_seqs == 1, (
|
||||
"Greedy sampling should have only one seq.")
|
||||
parent_ids = list(range(num_parent_seqs))
|
||||
next_token_ids = [samples[sample_idx]]
|
||||
next_token_ids = [samples_lst[sample_idx]]
|
||||
results.append((next_token_ids, parent_ids))
|
||||
sample_idx += num_parent_seqs
|
||||
return results
|
||||
@@ -394,7 +394,7 @@ def _beam_search_sample(
|
||||
next_token_ids = next_token_ids.tolist()
|
||||
else:
|
||||
# Generation phase.
|
||||
cumulative_logprobs: List[int] = [
|
||||
cumulative_logprobs: List[float] = [
|
||||
seq_group.seq_data[seq_id].cumulative_logprob
|
||||
for seq_id in seq_ids
|
||||
]
|
||||
@@ -466,8 +466,9 @@ def _sample_with_torch(
|
||||
categorized_seq_group_ids[sampling_type].append(i)
|
||||
|
||||
sample_results_dict: Dict[int, Tuple[List[int], List[int]]] = {}
|
||||
sample_metadata = {}
|
||||
multinomial_samples = {}
|
||||
sample_metadata: Dict[SamplingType,
|
||||
Tuple[List[int], List[SequenceGroupToSample]]] = {}
|
||||
multinomial_samples: Dict[SamplingType, torch.Tensor] = {}
|
||||
|
||||
# Create output tensor for sampled token ids.
|
||||
if include_gpu_probs_tensor:
|
||||
@@ -494,7 +495,7 @@ def _sample_with_torch(
|
||||
greedy_samples = torch.argmax(logprobs[long_sample_indices],
|
||||
dim=-1)
|
||||
|
||||
if include_gpu_probs_tensor:
|
||||
if sampled_token_ids_tensor is not None:
|
||||
# Store sampled tokens in output tensor.
|
||||
sampled_token_ids_tensor[
|
||||
long_sample_indices] = greedy_samples.unsqueeze(-1)
|
||||
@@ -522,7 +523,7 @@ def _sample_with_torch(
|
||||
probs[long_sample_indices], max_best_of_in_batch,
|
||||
**seeded_args)
|
||||
|
||||
if include_gpu_probs_tensor:
|
||||
if sampled_token_ids_tensor is not None:
|
||||
# Store sampled tokens in output tensor.
|
||||
sampled_token_ids_tensor[
|
||||
long_sample_indices] = multinomial_samples[sampling_type]
|
||||
@@ -571,7 +572,9 @@ def _sample_with_triton_kernel(
|
||||
categorized_seq_group_ids[sampling_type].append(i)
|
||||
|
||||
sample_results_dict: Dict[int, Tuple[List[int], List[int]]] = {}
|
||||
sample_metadata = {}
|
||||
sample_metadata: Dict[SamplingType,
|
||||
Tuple[List[int], List[SequenceGroupToSample],
|
||||
torch.Tensor, torch.Tensor]] = {}
|
||||
max_best_of_in_batch = 1
|
||||
|
||||
# Counterintiutively, having two loops here is actually faster.
|
||||
@@ -1008,14 +1011,14 @@ def _build_sampler_output(
|
||||
speculative decoding rejection sampling.
|
||||
"""
|
||||
|
||||
sampler_output = []
|
||||
sampler_output: List[CompletionSequenceGroupOutput] = []
|
||||
for (seq_group, sample_result, group_prompt_logprobs,
|
||||
group_sample_logprobs) in zip(sampling_metadata.seq_groups,
|
||||
sample_results, prompt_logprobs,
|
||||
sample_logprobs):
|
||||
seq_ids = seq_group.seq_ids
|
||||
next_token_ids, parent_ids = sample_result
|
||||
seq_outputs = []
|
||||
seq_outputs: List[SequenceOutput] = []
|
||||
for parent_id, next_token_id, logprobs in zip(parent_ids,
|
||||
next_token_ids,
|
||||
group_sample_logprobs):
|
||||
|
||||
Reference in New Issue
Block a user