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

fix: QWEN2MOE support for expert_feed_forward_length

Browse source 
previously, expert ff was taken from n_ff (intermediate size) but it is now properly taken from LLM_KV_EXPERT_FEED_FORWARD_LENGTH

n_ff_exp and n_ff_shexp are now properly calculated
This commit is contained in:
stefan 2024-06-14 11:38:12 +00:00
parent 06531cbaec
commit d9452267a0
4 changed files with 47 additions and 34 deletions
Download patch file Download diff file Expand all files Collapse all files

3
convert-hf-to-gguf.py
View file

@ -1630,6 +1630,9 @@ class Qwen2MoeModel(Model):
if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
logger.info(f"gguf: expert feed forward length = {moe_intermediate_size}")
if (shared_expert_intermediate_size := self.hparams.get('shared_expert_intermediate_size')) is not None:
self.gguf_writer.add_expert_shared_feed_forward_length(shared_expert_intermediate_size)
logger.info(f"gguf: expert shared feed forward length = {shared_expert_intermediate_size}")
_experts: list[dict[str, Tensor]] | None = None

1
gguf-py/gguf/constants.py
View file

@ -40,6 +40,7 @@ class Keys:
LEADING_DENSE_BLOCK_COUNT = "{arch}.leading_dense_block_count"
FEED_FORWARD_LENGTH = "{arch}.feed_forward_length"
EXPERT_FEED_FORWARD_LENGTH = "{arch}.expert_feed_forward_length"
EXPERT_SHARED_FEED_FORWARD_LENGTH = "{arch}.expert_shared_feed_forward_length"
USE_PARALLEL_RESIDUAL = "{arch}.use_parallel_residual"
TENSOR_DATA_LAYOUT = "{arch}.tensor_data_layout"
EXPERT_COUNT = "{arch}.expert_count"

3
gguf-py/gguf/gguf_writer.py
View file

@ -383,6 +383,9 @@ class GGUFWriter:
def add_expert_feed_forward_length(self, length: int) -> None:
self.add_uint32(Keys.LLM.EXPERT_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
def add_expert_shared_feed_forward_length(self, length: int) -> None:
self.add_uint32(Keys.LLM.EXPERT_SHARED_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
def add_parallel_residual(self, use: bool) -> None:
self.add_bool(Keys.LLM.USE_PARALLEL_RESIDUAL.format(arch=self.arch), use)

14
llama.cpp
View file

@ -282,6 +282,7 @@ enum llm_kv {
LLM_KV_LEADING_DENSE_BLOCK_COUNT,
LLM_KV_FEED_FORWARD_LENGTH,
LLM_KV_EXPERT_FEED_FORWARD_LENGTH,
LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH,
LLM_KV_USE_PARALLEL_RESIDUAL,
LLM_KV_TENSOR_DATA_LAYOUT,
LLM_KV_EXPERT_COUNT,
@ -367,6 +368,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_LEADING_DENSE_BLOCK_COUNT, "%s.leading_dense_block_count" },
{ LLM_KV_FEED_FORWARD_LENGTH, "%s.feed_forward_length" },
{ LLM_KV_EXPERT_FEED_FORWARD_LENGTH, "%s.expert_feed_forward_length" },
{ LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, "%s.expert_shared_feed_forward_length" },
{ LLM_KV_USE_PARALLEL_RESIDUAL, "%s.use_parallel_residual" },
{ LLM_KV_TENSOR_DATA_LAYOUT, "%s.tensor_data_layout" },
{ LLM_KV_EXPERT_COUNT, "%s.expert_count" },
@ -1840,6 +1842,7 @@ struct llama_hparams {
uint32_t n_lora_q = 0;
uint32_t n_lora_kv = 0;
uint32_t n_ff_exp = 0;
uint32_t n_ff_shexp = 0;
uint32_t n_expert_shared = 0;
float expert_weights_scale = 0.0;
@ -1888,6 +1891,7 @@ struct llama_hparams {
if (this->n_lora_q != other.n_lora_q) return true;
if (this->n_lora_kv != other.n_lora_kv) return true;
if (this->n_ff_exp != other.n_ff_exp) return true;
if (this->n_ff_shexp != other.n_ff_shexp) return true;
if (this->n_expert_shared != other.n_expert_shared) return true;
if (this->rope_finetuned != other.rope_finetuned) return true;
@ -4248,6 +4252,7 @@ static void llm_load_hparams(
case LLM_ARCH_QWEN2MOE:
{
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false);
ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
switch (hparams.n_layer) {
@ -5024,6 +5029,7 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
if (model.arch == LLM_ARCH_QWEN2MOE) {
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
LLAMA_LOG_INFO("%s: n_ff_shexp = %d\n", __func__, hparams.n_ff_shexp);
}
}
@ -5817,11 +5823,11 @@ static bool llm_load_tensors(
layer.ffn_up_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert});
// Shared expert branch
auto n_ff_shared_exp = hparams.n_ff_exp && hparams.n_expert_used ? hparams.n_ff_exp * hparams.n_expert_used : n_ff;
auto n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff;
layer.ffn_gate_inp_shexp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", i), {n_embd});
layer.ffn_gate_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shared_exp});
layer.ffn_down_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_shared_exp, n_embd});
layer.ffn_up_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, n_ff_shared_exp});
layer.ffn_gate_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp});
layer.ffn_down_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd});
layer.ffn_up_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, n_ff_shexp});
}
} break;
case LLM_ARCH_PHI2: