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

llama : add Mixtral support (#4406)

Browse source 
* convert : support Mixtral as LLAMA arch

* convert : fix n_ff typo

* llama : model loading

* ggml : sync latest ggml_mul_mat_id

* llama : update graph to support MoE

* llama : fix cur -> cur_expert

* llama : first working version

* llama : fix expert weighting in the FFN

* ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only)

* ggml : add n_as argument to ggml_mul_mat_id

* ggml : fix ggml_get_rows to take into account ne02 / ne11

* metal : add more general support for ggml_get_rows + tests

* llama : add basic support for offloading moe with CUDA

* metal : add/mul/div use general kernel when src1 not cont

* metal : reduce the kernel launches for ggml_mul_mat_id

* ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D

* ggml : update get_rows f16 and q

* cuda : support non-contiguous src1 in get_rows

* llama : offload missing ffn_moe_silu

* metal : fix ggml_get_rows to work with non-cont src1

* metal : add indirect mat-vec kernels for all quantization types

* llama : do not quantize expert gating tensors

* llama : add n_expert and n_expert_used to hparams + change quants

* test-backend-ops : add moe test

* cuda : fix get_rows when ncols is odd

* convert : determine n_ctx correctly

* metal : fix ggml_mul_mat_id for F32

* test-backend-ops : make experts more evenly probable (test_moe)

* test-backend-ops : cleanup, add moe test for batches

* test-backend-ops : add cpy from f32 -> all types test

* test-backend-ops : fix dequantize block offset

* llama : fix hard-coded number of experts

* test-backend-ops : simplify and disable slow tests to avoid CI timeout

* test-backend-ops : disable MOE test with thread sanitizer

* cuda : fix mul_mat_id with multi gpu

* convert : use 1e6 rope_freq_base for mixtral

* convert : fix style

* convert : support safetensors format

* gguf-py : bump version

* metal : add cpy f16 -> f32 kernel

* metal : fix binary ops for ne10 % 4 != 0

* test-backend-ops : add one more sum_rows test

* ggml : do not use BLAS with ggml_mul_mat_id

* convert-hf : support for mixtral-instruct (#4428)

* convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct

* convert : use sentencepiece tokenizer for Mixtral-instruct

* convert : make flake8 happy

* metal : fix soft_max kernels

ref: 1914017863

* metal : limit kernels to not use more than the allowed threads

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: Radek Pilar <github@mrkva.eu>
This commit is contained in:
slaren 2023-12-13 13:04:25 +01:00 • committed by GitHub
parent fecac45658
commit 799a1cb13b
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
14 changed files with 2370 additions and 395 deletions
Download patch file Download diff file Expand all files Collapse all files

16
gguf-py/gguf/constants.py
View file

@ -38,6 +38,8 @@ class Keys:
FEED_FORWARD_LENGTH = "{arch}.feed_forward_length"
USE_PARALLEL_RESIDUAL = "{arch}.use_parallel_residual"
TENSOR_DATA_LAYOUT = "{arch}.tensor_data_layout"
EXPERT_COUNT = "{arch}.expert_count"
EXPERT_USED_COUNT = "{arch}.expert_used_count"
class Attention:
HEAD_COUNT = "{arch}.attention.head_count"
@ -111,10 +113,14 @@ class MODEL_TENSOR(IntEnum):
ATTN_NORM = auto()
ATTN_NORM_2 = auto()
ATTN_ROT_EMBD = auto()
FFN_GATE_INP = auto()
FFN_NORM = auto()
FFN_GATE = auto()
FFN_DOWN = auto()
FFN_UP = auto()
FFN_NORM = auto()
FFN_GATE_EXP = auto()
FFN_DOWN_EXP = auto()
FFN_UP_EXP = auto()
ATTN_Q_NORM = auto()
ATTN_K_NORM = auto()
@ -154,10 +160,14 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.ATTN_ROT_EMBD: "blk.{bid}.attn_rot_embd",
MODEL_TENSOR.ATTN_Q_NORM: "blk.{bid}.attn_q_norm",
MODEL_TENSOR.ATTN_K_NORM: "blk.{bid}.attn_k_norm",
MODEL_TENSOR.FFN_GATE_INP: "blk.{bid}.ffn_gate_inp",
MODEL_TENSOR.FFN_NORM: "blk.{bid}.ffn_norm",
MODEL_TENSOR.FFN_GATE: "blk.{bid}.ffn_gate",
MODEL_TENSOR.FFN_DOWN: "blk.{bid}.ffn_down",
MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up",
MODEL_TENSOR.FFN_GATE_EXP: "blk.{bid}.ffn_gate.{xid}",
MODEL_TENSOR.FFN_DOWN_EXP: "blk.{bid}.ffn_down.{xid}",
MODEL_TENSOR.FFN_UP_EXP: "blk.{bid}.ffn_up.{xid}",
}
MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@ -172,10 +182,14 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.ATTN_ROT_EMBD,
MODEL_TENSOR.FFN_GATE_INP,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.FFN_GATE_EXP,
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
],
MODEL_ARCH.GPTNEOX: [
MODEL_TENSOR.TOKEN_EMBD,

6
gguf-py/gguf/gguf_writer.py
View file

@ -339,6 +339,12 @@ class GGUFWriter:
def add_clamp_kqv(self, value: float) -> None:
self.add_float32(Keys.Attention.CLAMP_KQV.format(arch=self.arch), value)
def add_expert_count(self, count: int) -> None:
self.add_uint32(Keys.LLM.EXPERT_COUNT.format(arch=self.arch), count)
def add_expert_used_count(self, count: int) -> None:
self.add_uint32(Keys.LLM.EXPERT_USED_COUNT.format(arch=self.arch), count)
def add_layer_norm_eps(self, value: float) -> None:
self.add_float32(Keys.Attention.LAYERNORM_EPS.format(arch=self.arch), value)

39
gguf-py/gguf/tensor_mapping.py
View file

@ -149,6 +149,11 @@ class TensorNameMap:
"model.layers.{bid}.ln2", # yi
),
MODEL_TENSOR.FFN_GATE_INP: (
"layers.{bid}.feed_forward.gate", # mixtral
"model.layers.{bid}.block_sparse_moe.gate", # mixtral
),
# Feed-forward up
MODEL_TENSOR.FFN_UP: (
"gpt_neox.layers.{bid}.mlp.dense_h_to_4h", # gptneox
@ -164,11 +169,21 @@ class TensorNameMap:
"transformer.h.{bid}.mlp.w1", # qwen
),
MODEL_TENSOR.FFN_UP_EXP: (
"layers.{bid}.feed_forward.experts.{xid}.w3", # mixtral
"model.layers.{bid}.block_sparse_moe.experts.{xid}.w3", # mixtral
),
# Feed-forward gate
MODEL_TENSOR.FFN_GATE: (
"model.layers.{bid}.mlp.gate_proj", # llama-hf refact
"layers.{bid}.feed_forward.w1", # llama-pth
"transformer.h.{bid}.mlp.w2", # qwen
"model.layers.{bid}.mlp.gate_proj", # llama-hf refact
"layers.{bid}.feed_forward.w1", # llama-pth
"transformer.h.{bid}.mlp.w2", # qwen
),
MODEL_TENSOR.FFN_GATE_EXP: (
"layers.{bid}.feed_forward.experts.{xid}.w1", # mixtral
"model.layers.{bid}.block_sparse_moe.experts.{xid}.w1", # mixtral
),
# Feed-forward down
@ -185,6 +200,11 @@ class TensorNameMap:
"language_model.encoder.layers.{bid}.mlp.dense_4h_to_h", # persimmon
),
MODEL_TENSOR.FFN_DOWN_EXP: (
"layers.{bid}.feed_forward.experts.{xid}.w2", # mixtral
"model.layers.{bid}.block_sparse_moe.experts.{xid}.w2", # mixtral
),
MODEL_TENSOR.ATTN_Q_NORM: (
"language_model.encoder.layers.{bid}.self_attention.q_layernorm",
),
@ -213,11 +233,14 @@ class TensorNameMap:
for tensor, keys in self.block_mappings_cfg.items():
if tensor not in MODEL_TENSORS[arch]:
continue
tensor_name = TENSOR_NAMES[tensor].format(bid = bid)
self.mapping[tensor_name] = (tensor, tensor_name)
for key in keys:
key = key.format(bid = bid)
self.mapping[key] = (tensor, tensor_name)
# TODO: make this configurable
n_experts = 8
for xid in range(n_experts):
tensor_name = TENSOR_NAMES[tensor].format(bid = bid, xid = xid)
self.mapping[tensor_name] = (tensor, tensor_name)
for key in keys:
key = key.format(bid = bid, xid = xid)
self.mapping[key] = (tensor, tensor_name)
def get_type_and_name(self, key: str, try_suffixes: Sequence[str] = ()) -> tuple[MODEL_TENSOR, str] | None:
result = self.mapping.get(key)

2
gguf-py/pyproject.toml
View file

@ -1,6 +1,6 @@
[tool.poetry]
name = "gguf"
version = "0.6.0"
version = "0.7.0"
description = "Read and write ML models in GGUF for GGML"
authors = ["GGML <ggml@ggml.ai>"]
packages = [