huggingface/text-generation-inference
5
0
Fork 0
mirror of https://github.com/huggingface/text-generation-inference.git synced 2025-05-07 18:32:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

Working version.

Browse Source
This commit is contained in:
Ubuntu 2023-05-09 08:35:59 +00:00
parent 57a6cbff82
commit a86e4bf713
3 changed files with 20 additions and 22 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
Dockerfile
View File

@ -165,6 +165,7 @@ COPY --from=builder /usr/src/target/release/text-generation-router /usr/local/bi
# Install launcher # Install launcher
COPY --from=builder /usr/src/target/release/text-generation-launcher /usr/local/bin/text-generation-launcher COPY --from=builder /usr/src/target/release/text-generation-launcher /usr/local/bin/text-generation-launcher
RUN apt update && apt install build-essential g++ -y
COPY proto proto COPY proto proto
COPY server/requirements.txt server/requirements.txt COPY server/requirements.txt server/requirements.txt
COPY server/pyproject.toml server/pyproject.toml COPY server/pyproject.toml server/pyproject.toml
@ -176,7 +177,6 @@ COPY server/Makefile-transformers server/Makefile-transformers
RUN cd server && \ RUN cd server && \
make gen-server && \ make gen-server && \
pip install ".[bnb, accelerate]" --no-cache-dir pip install ".[bnb, accelerate]" --no-cache-dir
RUN apt update && apt install build-essential g++ -y
# AWS Sagemaker compatbile image # AWS Sagemaker compatbile image
FROM base as sagemaker FROM base as sagemaker

39
server/text_generation_server/models/custom_modeling/flash_llama_modeling.py
View File

@ -38,6 +38,7 @@ from flash_attn.layers.rotary import RotaryEmbedding
# from safetensors.torch import load_file # from safetensors.torch import load_file
from safetensors import safe_open from safetensors import safe_open
from huggingface_hub import hf_hub_download from huggingface_hub import hf_hub_download
from loguru import logger
HAS_BITS_AND_BYTES = True HAS_BITS_AND_BYTES = True
try: try:
@ -197,29 +198,30 @@ class FastLinear(nn.Linear):
with safe_open(filename, framework="pt", device=f"cuda:{rank}") as f: with safe_open(filename, framework="pt", device=f"cuda:{rank}") as f:
if name == 'self_attn.query_key_value': if name == 'self_attn.query_key_value':
query_name = f'model.layers.{layer}.self_attn' query_name = f'model.layers.{layer}.self_attn'
self.qlinear.qweight[:, : self.out_features // 3] = get_slice(f, f"{query_name}.q_proj.qweight") N = self.out_features // 3
self.qlinear.qweight[:, self.out_features // 3:-self.out_features // 3] = get_slice(f, f"{query_name}.k_proj.qweight") self.qlinear.qweight[:, : N] = get_slice(f, f"{query_name}.q_proj.qweight")
self.qlinear.qweight[:,-self.out_features // 3: ] = get_slice(f, f"{query_name}.v_proj.qweight") self.qlinear.qweight[:, N:2 * N] = get_slice(f, f"{query_name}.k_proj.qweight")
self.qlinear.qweight[:, 2*N:] = get_slice(f, f"{query_name}.v_proj.qweight")
N = self.qlinear.qzeros.shape[1] self.qlinear.scales[:, : N] = get_slice(f, f"{query_name}.q_proj.scales")
self.qlinear.qzeros[:, : N // 3] = get_slice(f, f"{query_name}.q_proj.qzeros") self.qlinear.scales[:, N:2*N] = get_slice(f, f"{query_name}.k_proj.scales")
self.qlinear.qzeros[:, N // 3:-N // 3] = get_slice(f, f"{query_name}.k_proj.qzeros") self.qlinear.scales[:, 2*N:] = get_slice(f, f"{query_name}.v_proj.scales")
self.qlinear.qzeros[:,-N // 3: ] = get_slice(f, f"{query_name}.v_proj.qzeros")
N = self.qlinear.qzeros.shape[1] // 3
self.qlinear.qzeros[:, :N] = get_slice(f, f"{query_name}.q_proj.qzeros")
self.qlinear.qzeros[:, N:2*N] = get_slice(f, f"{query_name}.k_proj.qzeros")
self.qlinear.qzeros[:, 2*N:] = get_slice(f, f"{query_name}.v_proj.qzeros")
self.qlinear.scales[:, : self.out_features // 3] = get_slice(f, f"{query_name}.q_proj.scales")
self.qlinear.scales[:, self.out_features // 3:-self.out_features // 3] = get_slice(f, f"{query_name}.k_proj.scales")
self.qlinear.scales[:,-self.out_features // 3: ] = get_slice(f, f"{query_name}.v_proj.scales")
torch.testing.assert_close(f.get_tensor(f"{query_name}.q_proj.g_idx"), f.get_tensor(f"{query_name}.k_proj.g_idx")) torch.testing.assert_close(f.get_tensor(f"{query_name}.q_proj.g_idx"), f.get_tensor(f"{query_name}.k_proj.g_idx"))
torch.testing.assert_close(f.get_tensor(f"{query_name}.q_proj.g_idx"), f.get_tensor(f"{query_name}.v_proj.g_idx")) torch.testing.assert_close(f.get_tensor(f"{query_name}.q_proj.g_idx"), f.get_tensor(f"{query_name}.v_proj.g_idx"))
self.qlinear.g_idx[:] = f.get_tensor(f"{query_name}.q_proj.g_idx") self.qlinear.g_idx[:] = f.get_tensor(f"{query_name}.q_proj.g_idx")
elif name == "self_attn.o_proj": elif name == "self_attn.o_proj":
self.qlinear.qweight = f.get_tensor(f"model.layers.{layer}.self_attn.o_proj.qweight") self.qlinear.qweight = get_slice(f, f"model.layers.{layer}.self_attn.o_proj.qweight")
self.qlinear.qzeros = f.get_tensor(f"model.layers.{layer}.self_attn.o_proj.qzeros") self.qlinear.qzeros = f.get_tensor(f"model.layers.{layer}.self_attn.o_proj.qzeros")
self.qlinear.scales = f.get_tensor(f"model.layers.{layer}.self_attn.o_proj.scales") self.qlinear.scales = f.get_tensor(f"model.layers.{layer}.self_attn.o_proj.scales")
self.qlinear.g_idx[:] = get_slice(f, f"model.layers.{layer}.self_attn.o_proj.g_idx") self.qlinear.g_idx = get_slice(f, f"model.layers.{layer}.self_attn.o_proj.g_idx")
import ipdb;ipdb.set_trace()
elif name == "mlp.gate_up_proj": elif name == "mlp.gate_up_proj":
N = self.qlinear.qweight.shape[1] // 2 N = self.qlinear.qweight.shape[1] // 2
@ -233,21 +235,16 @@ class FastLinear(nn.Linear):
self.qlinear.g_idx[:] = f.get_tensor(f"model.layers.{layer}.mlp.gate_proj.g_idx") self.qlinear.g_idx[:] = f.get_tensor(f"model.layers.{layer}.mlp.gate_proj.g_idx")
N = self.qlinear.qzeros.shape[1] // 2 N = self.qlinear.qzeros.shape[1] // 2
self.qlinear.qzeros[:, N:] = get_slice(f, f"model.layers.{layer}.mlp.up_proj.qzeros")
self.qlinear.qzeros[:, :N] = get_slice(f, f"model.layers.{layer}.mlp.gate_proj.qzeros") self.qlinear.qzeros[:, :N] = get_slice(f, f"model.layers.{layer}.mlp.gate_proj.qzeros")
self.qlinear.qzeros[:, N:] = get_slice(f, f"model.layers.{layer}.mlp.up_proj.qzeros")
elif name == "mlp.down_proj": elif name == "mlp.down_proj":
self.qlinear.qweight = f.get_tensor(f"model.layers.{layer}.mlp.down_proj.qweight") self.qlinear.qweight = get_slice(f, f"model.layers.{layer}.mlp.down_proj.qweight")
self.qlinear.qzeros = f.get_tensor(f"model.layers.{layer}.mlp.down_proj.qzeros") self.qlinear.qzeros = f.get_tensor(f"model.layers.{layer}.mlp.down_proj.qzeros")
self.qlinear.scales = f.get_tensor(f"model.layers.{layer}.mlp.down_proj.scales") self.qlinear.scales = f.get_tensor(f"model.layers.{layer}.mlp.down_proj.scales")
self.qlinear.g_idx[:] = get_slice(f, f"model.layers.{layer}.mlp.down_proj.g_idx") self.qlinear.g_idx = get_slice(f, f"model.layers.{layer}.mlp.down_proj.g_idx")
else: else:
raise ValueError("Not handled") raise ValueError("Not handled")
print(layer, name)
if name == 'self_attn.query_key_value':
out = self.qlinear(torch.zeros((6, self.in_features)).cuda().half())
if name == "self_attn.o_proj":
out = self.qlinear(torch.zeros((6, self.in_features)).cuda().half())
# Delete reference to data # Delete reference to data
self.weight = None self.weight = None

1
server/text_generation_server/quant/quant_linear.py
View File

@ -372,6 +372,7 @@ class QuantLinear(nn.Module):
def forward(self, x): def forward(self, x):
out_shape = x.shape[:-1] + (self.outfeatures, ) out_shape = x.shape[:-1] + (self.outfeatures, )
assert x.shape[-1] == self.infeatures
out = QuantLinearFunction.apply(x.reshape(-1, x.shape[-1]), self.qweight, self.scales, self.qzeros, self.g_idx, self.bits, self.maxq) out = QuantLinearFunction.apply(x.reshape(-1, x.shape[-1]), self.qweight, self.scales, self.qzeros, self.g_idx, self.bits, self.maxq)
out = out + self.bias if self.bias is not None else out out = out + self.bias if self.bias is not None else out
return out.reshape(out_shape) return out.reshape(out_shape)