vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

llama : add support for BitnetForCausalLM (#7931)

Browse source 
* hf bitnet v1

* hf bitnet e2e v2

* finish bitnet e2e

* finish f16 hf bitnet e2e

* remove unsed

* finish bitnet i2 e2e

* move i2s to quantize v1

* move i2 to quantize

* clean code

* clean code 2

* fix codestyle

* fix code

* fix

* fix code

* fix merge

* remove unused

* change table name

* fix whitespace

* delete redundant

* i2_s to absmax

* finish i2_s/i8_s vec_dot x86 simd

* i2s->q22

* fix code

* remove block scale

* add dequantize

* fix seq

* update avx2

* remove q2_2

* remove q22_grid

* fix whitespace

* reuse llm_build_kv

* fix bo

---------

Co-authored-by: root <root@wangjinheng>
This commit is contained in:
Eddie-Wang 2024-06-24 02:27:57 +08:00 • committed by GitHub
parent 6a2f298bd7
commit e112b610a1
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 307 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

42
convert-hf-to-gguf.py
View file

@ -1404,6 +1404,48 @@ class LlamaModel(Model):
raise ValueError(f"Unprocessed experts: {experts}")
@Model.register("BitnetForCausalLM")
class BitnetModel(Model):
model_arch = gguf.MODEL_ARCH.BITNET
def set_vocab(self):
self._set_vocab_sentencepiece()
def set_gguf_parameters(self):
super().set_gguf_parameters()
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
self.gguf_writer.add_rope_scaling_factor(1.0)
def weight_quant(self, weight):
dtype = weight.dtype
weight = weight.float()
s = 1 / weight.abs().mean().clamp(min=1e-5)
weight = (weight * s).round().clamp(-1, 1) / s
scale = weight.abs().max().unsqueeze(0)
weight = torch.where(weight.abs().less(1e-6), 0, weight).type(dtype)
weight = torch.sign(weight).type(dtype)
return weight.type(dtype), scale.type(torch.float32)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
new_name = self.map_tensor_name(name)
if any(self.match_model_tensor_name(new_name, key, bid) for key in [
gguf.MODEL_TENSOR.ATTN_Q,
gguf.MODEL_TENSOR.ATTN_K,
gguf.MODEL_TENSOR.ATTN_V,
gguf.MODEL_TENSOR.ATTN_OUT,
gguf.MODEL_TENSOR.FFN_UP,
gguf.MODEL_TENSOR.FFN_DOWN,
gguf.MODEL_TENSOR.FFN_GATE,
]):
# transform weight into 1/0/-1 (in fp32)
weight_torch, scale_torch = self.weight_quant(data_torch)
yield (new_name, weight_torch)
yield (new_name.removesuffix(".weight") + ".scale", scale_torch)
else:
yield (new_name, data_torch)
@Model.register("GrokForCausalLM")
class GrokModel(Model):
model_arch = gguf.MODEL_ARCH.GROK