Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

fix(server): llama v2 GPTQ #648

Merged
merged 2 commits into from
Jul 20, 2023
Merged

fix(server): llama v2 GPTQ #648

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
5882768
nit
fxmarty Jul 19, 2023
2080735
taste
fxmarty Jul 19, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
35 changes: 19 additions & 16 deletions server/text_generation_server/models/custom_modeling/flash_llama_modeling.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -148,24 +148,27 @@ def forward(self, hidden_states, residual=None):


def _load_gqa(config, prefix: str, weights):
w = [
weights.get_sharded(f"{prefix}.q_proj.weight", dim=0),
weights.get_sharded(f"{prefix}.k_proj.weight", dim=0),
weights.get_sharded(f"{prefix}.v_proj.weight", dim=0),
]
weight = torch.cat(w, dim=0)
weight = weight.to(dtype=weights.dtype).to(device=weights.device)
bias = None
assert config.hidden_size % config.num_attention_heads == 0
head_size = config.hidden_size // config.num_attention_heads
assert config.num_attention_heads % weights.process_group.size() == 0
num_heads = config.num_attention_heads // weights.process_group.size()
num_key_value_heads = config.num_key_value_heads // weights.process_group.size()
assert list(weight.shape) == [
(num_heads + 2 * num_key_value_heads) * head_size,
config.hidden_size,
], f"{list(weight.shape)} != {[(num_heads + 2 * config.num_key_value_heads) * head_size, config.hidden_size]}"
return TensorParallelColumnLinear(get_linear(weight, bias, config.quantize))

weight = weights.get_multi_weights_col(
prefixes=[f"{prefix}.q_proj", f"{prefix}.k_proj", f"{prefix}.v_proj"],
quantize=config.quantize,
dim=0
)

if config.quantize != "gptq":
weight = weight.to(dtype=weights.dtype).to(device=weights.device)

head_size = config.hidden_size // config.num_attention_heads
num_heads = config.num_attention_heads // weights.process_group.size()
num_key_value_heads = config.num_key_value_heads // weights.process_group.size()
assert list(weight.shape) == [
(num_heads + 2 * num_key_value_heads) * head_size,
config.hidden_size,
], f"{list(weight.shape)} != {[(num_heads + 2 * config.num_key_value_heads) * head_size, config.hidden_size]}"

return TensorParallelColumnLinear(get_linear(weight, bias=None, quantize=config.quantize))


class FlashLlamaAttention(torch.nn.Module):
Expand Down
Loading