llama: Add attention and final logit soft-capping, update scaling factor to Gemma2 (#8197)

* Add attention and final logit softcapping.

* fix

* Add custom add_ functions

* Disable flash attention for Gemma2

* Update src/llama.cpp

Co-authored-by: slaren <slarengh@gmail.com>

* Add default value for attention and final logit softcap value

* Add custom kq scaling from Gemma2Attention

* Remove custom pre attention scaling and use computed value instead.

---------

Co-authored-by: slaren <slarengh@gmail.com>
This commit is contained in:
Andrei 2024-06-29 20:44:08 -07:00 committed by GitHub
parent 72272b83a3
commit 1c5eba6f8e
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
4 changed files with 46 additions and 3 deletions

View File

@ -2363,6 +2363,12 @@ class Gemma2Model(Model):
self.gguf_writer.add_key_length(hparams["head_dim"]) self.gguf_writer.add_key_length(hparams["head_dim"])
self.gguf_writer.add_value_length(hparams["head_dim"]) self.gguf_writer.add_value_length(hparams["head_dim"])
self.gguf_writer.add_file_type(self.ftype) self.gguf_writer.add_file_type(self.ftype)
self.gguf_writer.add_attn_logit_softcapping(
self.hparams["attn_logit_softcapping"]
)
self.gguf_writer.add_final_logit_softcapping(
self.hparams["final_logit_softcapping"]
)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
del bid # unusem del bid # unusem

View File

@ -50,6 +50,8 @@ class Keys:
POOLING_TYPE = "{arch}.pooling_type" POOLING_TYPE = "{arch}.pooling_type"
LOGIT_SCALE = "{arch}.logit_scale" LOGIT_SCALE = "{arch}.logit_scale"
DECODER_START_TOKEN_ID = "{arch}.decoder_start_token_id" DECODER_START_TOKEN_ID = "{arch}.decoder_start_token_id"
ATTN_LOGIT_SOFTCAPPING = "{arch}.attn_logit_softcapping"
FINAL_LOGIT_SOFTCAPPING = "{arch}.final_logit_softcapping"
class Attention: class Attention:
HEAD_COUNT = "{arch}.attention.head_count" HEAD_COUNT = "{arch}.attention.head_count"

View File

@ -516,6 +516,12 @@ class GGUFWriter:
def add_logit_scale(self, value: float) -> None: def add_logit_scale(self, value: float) -> None:
self.add_float32(Keys.LLM.LOGIT_SCALE.format(arch=self.arch), value) self.add_float32(Keys.LLM.LOGIT_SCALE.format(arch=self.arch), value)
def add_attn_logit_softcapping(self, value: float) -> None:
self.add_float32(Keys.LLM.ATTN_LOGIT_SOFTCAPPING.format(arch=self.arch), value)
def add_final_logit_softcapping(self, value: float) -> None:
self.add_float32(Keys.LLM.FINAL_LOGIT_SOFTCAPPING.format(arch=self.arch), value)
def add_expert_count(self, count: int) -> None: def add_expert_count(self, count: int) -> None:
self.add_uint32(Keys.LLM.EXPERT_COUNT.format(arch=self.arch), count) self.add_uint32(Keys.LLM.EXPERT_COUNT.format(arch=self.arch), count)

View File

@ -302,6 +302,8 @@ enum llm_kv {
LLM_KV_POOLING_TYPE, LLM_KV_POOLING_TYPE,
LLM_KV_LOGIT_SCALE, LLM_KV_LOGIT_SCALE,
LLM_KV_DECODER_START_TOKEN_ID, LLM_KV_DECODER_START_TOKEN_ID,
LLM_KV_ATTN_LOGIT_SOFTCAPPING,
LLM_KV_FINAL_LOGIT_SOFTCAPPING,
LLM_KV_ATTENTION_HEAD_COUNT, LLM_KV_ATTENTION_HEAD_COUNT,
LLM_KV_ATTENTION_HEAD_COUNT_KV, LLM_KV_ATTENTION_HEAD_COUNT_KV,
@ -392,6 +394,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_POOLING_TYPE , "%s.pooling_type" }, { LLM_KV_POOLING_TYPE , "%s.pooling_type" },
{ LLM_KV_LOGIT_SCALE, "%s.logit_scale" }, { LLM_KV_LOGIT_SCALE, "%s.logit_scale" },
{ LLM_KV_DECODER_START_TOKEN_ID, "%s.decoder_start_token_id" }, { LLM_KV_DECODER_START_TOKEN_ID, "%s.decoder_start_token_id" },
{ LLM_KV_ATTN_LOGIT_SOFTCAPPING, "%s.attn_logit_softcapping" },
{ LLM_KV_FINAL_LOGIT_SOFTCAPPING, "%s.final_logit_softcapping" },
{ LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" }, { LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" },
{ LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" }, { LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" },
@ -2099,6 +2103,9 @@ struct llama_hparams {
float f_norm_eps; float f_norm_eps;
float f_norm_rms_eps; float f_norm_rms_eps;
float f_attn_logit_softcapping = 50.0f;
float f_final_logit_softcapping = 30.0f;
float rope_attn_factor = 1.0f; float rope_attn_factor = 1.0f;
float rope_freq_base_train; float rope_freq_base_train;
float rope_freq_scale_train; float rope_freq_scale_train;
@ -2117,6 +2124,7 @@ struct llama_hparams {
bool causal_attn = true; bool causal_attn = true;
bool use_alibi = false; bool use_alibi = false;
bool attn_soft_cap = false;
enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_NONE; enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_NONE;
enum llama_rope_type rope_type = LLAMA_ROPE_TYPE_NONE; enum llama_rope_type rope_type = LLAMA_ROPE_TYPE_NONE;
@ -4702,6 +4710,9 @@ static void llm_load_hparams(
case LLM_ARCH_GEMMA2: case LLM_ARCH_GEMMA2:
{ {
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
ml.get_key(LLM_KV_ATTN_LOGIT_SOFTCAPPING, hparams.f_attn_logit_softcapping, false);
ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING, hparams.f_final_logit_softcapping, false);
hparams.attn_soft_cap = true;
switch (hparams.n_layer) { switch (hparams.n_layer) {
case 42: model.type = e_model::MODEL_9B; break; case 42: model.type = e_model::MODEL_9B; break;
@ -7579,6 +7590,12 @@ static struct ggml_tensor * llm_build_kqv(
kq = ggml_scale(ctx, kq, 30); kq = ggml_scale(ctx, kq, 30);
} }
if (hparams.attn_soft_cap) {
kq = ggml_scale(ctx, kq, 1.0f / hparams.f_attn_logit_softcapping);
kq = ggml_tanh(ctx, kq);
kq = ggml_scale(ctx, kq, hparams.f_attn_logit_softcapping);
}
kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_scale, hparams.f_max_alibi_bias); kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_scale, hparams.f_max_alibi_bias);
cb(kq, "kq_soft_max_ext", il); cb(kq, "kq_soft_max_ext", il);
@ -11039,7 +11056,7 @@ struct llm_build_context {
ext_factor, attn_factor, beta_fast, beta_slow); ext_factor, attn_factor, beta_fast, beta_slow);
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k))); Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd / n_head)));
cb(Qcur, "Qcur_scaled", il); cb(Qcur, "Qcur_scaled", il);
Kcur = ggml_rope_ext( Kcur = ggml_rope_ext(
@ -11106,6 +11123,12 @@ struct llm_build_context {
// lm_head // lm_head
cur = ggml_mul_mat(ctx0, model.output, cur); cur = ggml_mul_mat(ctx0, model.output, cur);
// final logit soft-capping
cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
cur = ggml_tanh(ctx0, cur);
cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
cb(cur, "result_output", -1); cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur); ggml_build_forward_expand(gf, cur);
@ -17379,6 +17402,12 @@ struct llama_context * llama_new_context_with_model(
params.flash_attn = false; params.flash_attn = false;
} }
if (params.flash_attn && model->hparams.attn_soft_cap) {
LLAMA_LOG_WARN("%s: flash_attn is not compatible with attn_soft_cap - forcing off\n", __func__);
params.flash_attn = false;
}
if (params.flash_attn && model->hparams.n_embd_head_k != model->hparams.n_embd_head_v) { if (params.flash_attn && model->hparams.n_embd_head_k != model->hparams.n_embd_head_v) {
LLAMA_LOG_WARN("%s: flash_attn requires n_embd_head_k == n_embd_head_v - forcing off\n", __func__); LLAMA_LOG_WARN("%s: flash_attn requires n_embd_head_k == n_embd_head_v - forcing off\n", __func__);
params.flash_attn = false; params.flash_attn = false;