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

ggml : reduce hash table reset cost (#8698)

Browse source 
* ggml : reduce hash table reset cost

* fix unreachable code warnings after GGML_ASSERT(false)

* GGML_ASSERT(false) -> GGML_ABORT("fatal error")

* GGML_ABORT use format string
This commit is contained in:
slaren 2024-07-27 04:41:55 +02:00 • committed by GitHub
parent 01245f5b16
commit 2b1f616b20
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
46 changed files with 851 additions and 754 deletions
Download patch file Download diff file Expand all files Collapse all files

4
src/llama-grammar.cpp
View file

@ -221,7 +221,7 @@ static void llama_grammar_advance_stack(
// end of alternate (LLAMA_GRETYPE_END, LLAMA_GRETYPE_ALT) or middle of char range
// (LLAMA_GRETYPE_CHAR_ALT, LLAMA_GRETYPE_CHAR_RNG_UPPER); stack should never be left on
// those
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
}
@ -517,7 +517,7 @@ void llama_grammar_accept_token_impl(struct llama_grammar * grammar, const struc
return;
}
}
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
const std::string & piece = vocab->cache_token_to_piece.at(token);

14
src/llama-vocab.cpp
View file

@ -152,14 +152,14 @@ static uint8_t llama_token_to_byte(const llama_vocab & vocab, llama_token id) {
return strtol(buf.c_str(), NULL, 16);
}
case LLAMA_VOCAB_TYPE_BPE: {
GGML_ASSERT(false);
return unicode_utf8_to_byte(token_data.text); // TODO: why is this here after GGML_ASSERT?
GGML_ABORT("fatal error");
//return unicode_utf8_to_byte(token_data.text); // TODO: why is this here after GGML_ASSERT?
}
case LLAMA_VOCAB_TYPE_WPM: {
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
default:
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
}
@ -1396,7 +1396,7 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
}
} break;
case LLAMA_VOCAB_TYPE_NONE:
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
return output;
@ -1422,7 +1422,7 @@ llama_token llama_byte_to_token_impl(const llama_vocab & vocab, uint8_t ch) {
return vocab.token_to_id.at(unicode_byte_to_utf8(ch));
}
default:
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
}
@ -1606,7 +1606,7 @@ int32_t llama_token_to_piece_impl(const struct llama_vocab & vocab, llama_token
break;
}
default:
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
}

36
src/llama.cpp
View file

@ -2259,8 +2259,7 @@ struct llama_hparams {
return n_head_arr[il];
}
GGML_ASSERT(false);
return 0;
GGML_ABORT("fatal error");
}
uint32_t n_head_kv(uint32_t il = 0) const {
@ -2268,8 +2267,7 @@ struct llama_hparams {
return n_head_kv_arr[il];
}
GGML_ASSERT(false);
return 0;
GGML_ABORT("fatal error");
}
uint32_t n_ff(uint32_t il = 0) const {
@ -2277,8 +2275,7 @@ struct llama_hparams {
return n_ff_arr[il];
}
GGML_ASSERT(false);
return 0;
GGML_ABORT("fatal error");
}
uint32_t n_gqa(uint32_t il = 0) const {
@ -8072,7 +8069,7 @@ static struct ggml_tensor * llm_build_moe_ffn(
cb(gate, "ffn_moe_gelu", il);
} break;
default:
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
ggml_tensor * par = ggml_mul(ctx, up, gate); // [n_ff, n_expert_used, n_tokens]
@ -8635,8 +8632,8 @@ struct llm_build_context {
} break;
default:
{
GGML_ASSERT(false && "unknown pooling type");
} break;
GGML_ABORT("unknown pooling type");
}
}
cb(cur, "result_embd_pooled", -1);
@ -8891,7 +8888,7 @@ struct llm_build_context {
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd/n_head, n_head, n_tokens);
break;
default:
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
@ -11723,7 +11720,7 @@ struct llm_build_context {
switch (model.type) {
case e_model::MODEL_9B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k))); break;
case e_model::MODEL_27B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd / n_head))); break;
default: GGML_ASSERT(false);
default: GGML_ABORT("fatal error");
};
cb(Qcur, "Qcur_scaled", il);
@ -13888,7 +13885,7 @@ static struct ggml_cgraph * llama_build_graph(
result = llm.build_jais();
} break;
default:
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
// add on pooling layer
@ -14687,8 +14684,8 @@ static int llama_decode_internal(
} break;
case LLAMA_POOLING_TYPE_UNSPECIFIED:
{
GGML_ASSERT(false && "unknown pooling type");
} break;
GGML_ABORT("unknown pooling type");
}
}
}
n_outputs_prev += lctx.n_outputs;
@ -15079,7 +15076,7 @@ static void llama_kv_cache_update_internal(struct llama_context & lctx) {
// apply K-shift if needed
if (lctx.model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE && lctx.kv_self.has_shift) {
if (lctx.model.arch == LLM_ARCH_DEEPSEEK2) { // not supported due to MLA
GGML_ASSERT(false && "Deepseek2 does not support K-shift");
GGML_ABORT("Deepseek2 does not support K-shift");
}
{
@ -15218,7 +15215,7 @@ static void llama_tensor_dequantize_internal(
} else if (ggml_is_quantized(tensor->type)) {
qtype.to_float(tensor->data, f32_output, nelements);
} else {
GGML_ASSERT(false); // unreachable
GGML_ABORT("fatal error"); // unreachable
}
return;
}
@ -16904,8 +16901,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
// all model arches should be listed explicitly here
case LLM_ARCH_UNKNOWN:
GGML_ASSERT(false && "unknown architecture");
break;
GGML_ABORT("unknown architecture");
}
return LLAMA_ROPE_TYPE_NONE;
@ -18469,7 +18465,7 @@ float * llama_get_logits_ith(struct llama_context * ctx, int32_t i) {
} catch (const std::exception & err) {
LLAMA_LOG_ERROR("%s: invalid logits id %d, reason: %s\n", __func__, i, err.what());
#ifndef NDEBUG
GGML_ASSERT(false);
GGML_ABORT("fatal error");
#endif
return nullptr;
}
@ -18514,7 +18510,7 @@ float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i) {
} catch (const std::exception & err) {
LLAMA_LOG_ERROR("%s: invalid embeddings id %d, reason: %s\n", __func__, i, err.what());
#ifndef NDEBUG
GGML_ASSERT(false);
GGML_ABORT("fatal error");
#endif
return nullptr;
}