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

llama : quantize more Mamba tensors

Browse source 
* llama : use f16 as the fallback of fallback quant types
This commit is contained in:
Francis Couture-Harpin 2024-08-19 12:44:35 -04:00
parent 5c0f108e15
commit 3491291a32
2 changed files with 4 additions and 19 deletions
Download patch file Download diff file Expand all files Collapse all files

18
convert_hf_to_gguf.py
View file

@ -295,6 +295,7 @@ class Model:
gguf.MODEL_TENSOR.FFN_GATE_INP, gguf.MODEL_TENSOR.FFN_GATE_INP,
gguf.MODEL_TENSOR.POS_EMBD, gguf.MODEL_TENSOR.POS_EMBD,
gguf.MODEL_TENSOR.TOKEN_TYPES, gguf.MODEL_TENSOR.TOKEN_TYPES,
gguf.MODEL_TENSOR.SSM_CONV1D,
) )
) )
or not name.endswith(".weight") or not name.endswith(".weight")
@ -2786,23 +2787,6 @@ class MambaModel(Model):
return [(new_name, data_torch)] return [(new_name, data_torch)]
def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
if bid is not None and new_name in (
self.format_tensor_name(
n, bid, ".weight" if name.endswith(".weight") else ""
)
for n in [
gguf.MODEL_TENSOR.SSM_CONV1D,
gguf.MODEL_TENSOR.SSM_X,
gguf.MODEL_TENSOR.SSM_DT,
gguf.MODEL_TENSOR.SSM_A,
gguf.MODEL_TENSOR.SSM_D,
]
):
return gguf.GGMLQuantizationType.F32
return super().tensor_force_quant(name, new_name, bid, n_dims)
@Model.register("CohereForCausalLM") @Model.register("CohereForCausalLM")
class CommandR2Model(Model): class CommandR2Model(Model):

5
src/llama.cpp
View file

@ -16122,6 +16122,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
case GGML_TYPE_Q6_K: new_type = GGML_TYPE_Q8_0; break; case GGML_TYPE_Q6_K: new_type = GGML_TYPE_Q8_0; break;
default: throw std::runtime_error("\nUnsupported tensor size encountered\n"); default: throw std::runtime_error("\nUnsupported tensor size encountered\n");
} }
if (tensor->ne[0] % ggml_blck_size(new_type) != 0) {
new_type = GGML_TYPE_F16;
}
LLAMA_LOG_WARN(" - using fallback quantization %s\n", ggml_type_name(new_type)); LLAMA_LOG_WARN(" - using fallback quantization %s\n", ggml_type_name(new_type));
++qs.n_fallback; ++qs.n_fallback;
} }
@ -16450,8 +16453,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
// do not quantize Mamba's small yet 2D weights // do not quantize Mamba's small yet 2D weights
// NOTE: can't use LLM_TN here because the layer number is not known // NOTE: can't use LLM_TN here because the layer number is not known
quantize &= name.find("ssm_conv1d.weight") == std::string::npos; quantize &= name.find("ssm_conv1d.weight") == std::string::npos;
quantize &= name.find("ssm_x.weight") == std::string::npos;
quantize &= name.find("ssm_dt.weight") == std::string::npos;
// do not quantize relative position bias (T5) // do not quantize relative position bias (T5)
quantize &= name.find("attn_rel_b.weight") == std::string::npos; quantize &= name.find("attn_rel_b.weight") == std::string::npos;