biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

TP/quantization/weight loading refactor part 1 - Simplify parallel linear logic (#1181)

Browse Source
This commit is contained in:
Zhuohan Li
2023-10-02 15:36:09 -07:00
committed by GitHub
parent 84e4e37d14
commit ba0bfd40e2
42 changed files with 819 additions and 1547 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
tests/kernels/test_cache.py
View File

@@ -106,14 +106,14 @@ def test_reshape_and_cache(
# Create a random slot mapping.
num_slots = block_size * num_blocks
slot_mapping = random.sample(range(num_slots), num_tokens)
slot_mapping = torch.tensor(slot_mapping, dtype=torch.int, device='cuda')
slot_mapping = torch.tensor(slot_mapping, dtype=torch.int, device="cuda")
qkv = torch.randn(num_tokens,
3,
num_heads,
head_size,
dtype=dtype,
device='cuda')
device="cuda")
_, key, value = qkv.unbind(dim=1)
# Create the KV caches.
@@ -132,7 +132,7 @@ def test_reshape_and_cache(
# Run the reference implementation.
reshaped_key = key.reshape(num_tokens, *key_cache[0, :, :, 0, :].shape)
block_indicies = torch.div(slot_mapping, block_size, rounding_mode='floor')
block_indicies = torch.div(slot_mapping, block_size, rounding_mode="floor")
block_indicies = block_indicies.cpu().tolist()
block_offsets = slot_mapping % block_size
block_offsets = block_offsets.cpu().tolist()