Skip to content

Commit

Permalink
update: support Qwen2-57B-A14B (#7835)
Browse files Browse the repository at this point in the history 
* update: convert-hf-to-gguf.py to support Qwen2-57B-A14B

* fix: QWEN2MOE support for expert_feed_forward_length

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_shared_exp are now properly calculated

* update: convert-hf-to-gguf.py cleanup for Qwen2MoeForCausalLM

* fix: QWEN2MOE support for expert_feed_forward_length

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
  • Loading branch information
@legraphista
legraphista authored Jun 17, 2024
1 parent 5b6da18 commit a94e6ff
Show file tree
Hide file tree
Showing 4 changed files with 57 additions and 34 deletions.
6 changes: 6 additions & 0 deletions convert-hf-to-gguf.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1632,6 +1632,12 @@ def set_gguf_parameters(self):
super().set_gguf_parameters()
if (n_experts := self.hparams.get("num_experts")) is not None:
self.gguf_writer.add_expert_count(n_experts)
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

Expand Down
31 changes: 16 additions & 15 deletions gguf-py/gguf/constants.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -33,21 +33,22 @@ class General:
FILE_TYPE = "general.file_type"

class LLM:
VOCAB_SIZE = "{arch}.vocab_size"
CONTEXT_LENGTH = "{arch}.context_length"
EMBEDDING_LENGTH = "{arch}.embedding_length"
BLOCK_COUNT = "{arch}.block_count"
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"
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"
EXPERT_SHARED_COUNT = "{arch}.expert_shared_count"
EXPERT_WEIGHTS_SCALE = "{arch}.expert_weights_scale"
POOLING_TYPE = "{arch}.pooling_type"
LOGIT_SCALE = "{arch}.logit_scale"
VOCAB_SIZE = "{arch}.vocab_size"
CONTEXT_LENGTH = "{arch}.context_length"
EMBEDDING_LENGTH = "{arch}.embedding_length"
BLOCK_COUNT = "{arch}.block_count"
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"
EXPERT_USED_COUNT = "{arch}.expert_used_count"
EXPERT_SHARED_COUNT = "{arch}.expert_shared_count"
EXPERT_WEIGHTS_SCALE = "{arch}.expert_weights_scale"
POOLING_TYPE = "{arch}.pooling_type"
LOGIT_SCALE = "{arch}.logit_scale"

class Attention:
HEAD_COUNT = "{arch}.attention.head_count"
Expand Down
3 changes: 3 additions & 0 deletions gguf-py/gguf/gguf_writer.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -394,6 +394,9 @@ def add_feed_forward_length(self, length: int) -> None:
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)

Expand Down
51 changes: 32 additions & 19 deletions llama.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -286,6 +286,7 @@ 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,
Expand Down Expand Up @@ -364,21 +365,22 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_GENERAL_SOURCE_URL, "general.source.url" },
{ LLM_KV_GENERAL_SOURCE_HF_REPO, "general.source.huggingface.repository" },

{ LLM_KV_VOCAB_SIZE, "%s.vocab_size" },
{ LLM_KV_CONTEXT_LENGTH, "%s.context_length" },
{ LLM_KV_EMBEDDING_LENGTH, "%s.embedding_length" },
{ LLM_KV_BLOCK_COUNT, "%s.block_count" },
{ 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_USE_PARALLEL_RESIDUAL, "%s.use_parallel_residual" },
{ LLM_KV_TENSOR_DATA_LAYOUT, "%s.tensor_data_layout" },
{ LLM_KV_EXPERT_COUNT, "%s.expert_count" },
{ LLM_KV_EXPERT_USED_COUNT, "%s.expert_used_count" },
{ LLM_KV_EXPERT_SHARED_COUNT, "%s.expert_shared_count" },
{ LLM_KV_EXPERT_WEIGHTS_SCALE, "%s.expert_weights_scale" },
{ LLM_KV_POOLING_TYPE , "%s.pooling_type" },
{ LLM_KV_LOGIT_SCALE, "%s.logit_scale" },
{ LLM_KV_VOCAB_SIZE, "%s.vocab_size" },
{ LLM_KV_CONTEXT_LENGTH, "%s.context_length" },
{ LLM_KV_EMBEDDING_LENGTH, "%s.embedding_length" },
{ LLM_KV_BLOCK_COUNT, "%s.block_count" },
{ 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" },
{ LLM_KV_EXPERT_USED_COUNT, "%s.expert_used_count" },
{ LLM_KV_EXPERT_SHARED_COUNT, "%s.expert_shared_count" },
{ LLM_KV_EXPERT_WEIGHTS_SCALE, "%s.expert_weights_scale" },
{ LLM_KV_POOLING_TYPE , "%s.pooling_type" },
{ LLM_KV_LOGIT_SCALE, "%s.logit_scale" },

{ LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" },
{ LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" },
Expand Down Expand Up @@ -1970,6 +1972,7 @@ 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;

Expand Down Expand Up @@ -2018,6 +2021,7 @@ 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;
Expand Down Expand Up @@ -4455,6 +4459,9 @@ static void llm_load_hparams(
} break;
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) {
case 24: model.type = e_model::MODEL_A2_7B; break;
Expand Down Expand Up @@ -5240,6 +5247,11 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale);
LLAMA_LOG_INFO("%s: rope_yarn_log_mul = %.4f\n", __func__, hparams.rope_yarn_log_mul);
}

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);
}
}

// Returns false if cancelled by progress_callback
Expand Down Expand Up @@ -6026,16 +6038,17 @@ static bool llm_load_tensors(
GGML_ASSERT(hparams.n_expert_used > 0);

// MoE branch
auto n_ff_exp = n_ff / hparams.n_expert_used;
auto n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / hparams.n_expert_used;
layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert});
layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert});
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_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});
layer.ffn_down_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff, n_embd});
layer.ffn_up_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, n_ff});
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:
Expand Down

0 comments on commit a94e6ff

Please sign in to comment.