From ba2135ccae7462470b3865c6e41d2e1d734eac05 Mon Sep 17 00:00:00 2001 From: slaren Date: Wed, 21 Feb 2024 22:18:23 +0100 Subject: [PATCH 01/13] gemma : allow offloading the output tensor (#5646) --- llama.cpp | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/llama.cpp b/llama.cpp index 3a226c426..4054d5da6 100644 --- a/llama.cpp +++ b/llama.cpp @@ -4394,6 +4394,8 @@ static bool llm_load_tensors( // output model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // same as tok_embd, duplicated to allow offloading + ml.n_created--; // artificial tensor const int64_t n_ff = hparams.n_ff; const int64_t n_embd_head_k = hparams.n_embd_head_k; @@ -7525,7 +7527,7 @@ struct llm_build_context { cb(cur, "result_norm", -1); // lm_head - cur = ggml_mul_mat(ctx0, model.tok_embd, cur); + cur = ggml_mul_mat(ctx0, model.output, cur); cb(cur, "result_output", -1); ggml_build_forward_expand(gf, cur); From 7fe4678b0244ba7b03eae66ebeaa947e2770bb1a Mon Sep 17 00:00:00 2001 From: slaren Date: Wed, 21 Feb 2024 22:52:39 +0100 Subject: [PATCH 02/13] llama : fix session save/load with quantized KV (#5649) --- llama.cpp | 20 ++++++++++---------- 1 file changed, 10 insertions(+), 10 deletions(-) diff --git a/llama.cpp b/llama.cpp index 4054d5da6..d763cc80c 100644 --- a/llama.cpp +++ b/llama.cpp @@ -12176,18 +12176,19 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat data_ctx->write(&kv_used, sizeof(kv_used)); if (kv_buf_size) { - const size_t elt_size = ggml_element_size(kv_self.k_l[0]); - std::vector tmp_buf; for (int il = 0; il < (int) n_layer; ++il) { - tmp_buf.resize(elt_size*n_embd_k_gqa*kv_head); + size_t k_size = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*kv_head); + tmp_buf.resize(k_size); ggml_backend_tensor_get(kv_self.k_l[il], tmp_buf.data(), 0, tmp_buf.size()); data_ctx->write(tmp_buf.data(), tmp_buf.size()); // v is not contiguous, copy row by row - tmp_buf.resize(elt_size*kv_head); + size_t v_row_size = ggml_row_size(kv_self.v_l[il]->type, kv_head); + size_t v_row_stride = ggml_row_size(kv_self.v_l[il]->type, n_ctx); + tmp_buf.resize(v_row_size); for (int ir = 0; ir < (int) n_embd_v_gqa; ++ir) { - ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), ir*elt_size*n_ctx, tmp_buf.size()); + ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), ir*v_row_stride, tmp_buf.size()); data_ctx->write(tmp_buf.data(), tmp_buf.size()); } } @@ -12289,17 +12290,16 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) { if (kv_buf_size) { GGML_ASSERT(kv_self.total_size() == kv_buf_size); - const size_t elt_size = ggml_element_size(kv_self.k_l[0]); - for (int il = 0; il < (int) n_layer; ++il) { - size_t k_size = elt_size*n_embd_k_gqa*kv_head; + size_t k_size = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*kv_head); ggml_backend_tensor_set(kv_self.k_l[il], inp, 0, k_size); inp += k_size; // v is not contiguous, copy row by row - size_t v_row_size = elt_size*kv_head; + size_t v_row_size = ggml_row_size(kv_self.v_l[il]->type, kv_head); + size_t v_row_stride = ggml_row_size(kv_self.v_l[il]->type, n_ctx); for (int ir = 0; ir < (int) n_embd_v_gqa; ++ir) { - ggml_backend_tensor_set(kv_self.v_l[il], inp, ir*elt_size*n_ctx, v_row_size); + ggml_backend_tensor_set(kv_self.v_l[il], inp, ir*v_row_stride, v_row_size); inp += v_row_size; } } From 7c8bcc11dc61cf5930b70cd0168b84afcebe12a9 Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Thu, 22 Feb 2024 00:31:00 +0100 Subject: [PATCH 03/13] Add docs for llama_chat_apply_template (#5645) * add docs for llama_chat_apply_template * fix typo --- examples/server/README.md | 1 + llama.h | 2 +- 2 files changed, 2 insertions(+), 1 deletion(-) diff --git a/examples/server/README.md b/examples/server/README.md index 6d9f96cd4..4b24ee5dc 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -41,6 +41,7 @@ see https://github.com/ggerganov/llama.cpp/issues/1437 - `--grp-attn-w`: Set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n` - `-n, --n-predict`: Set the maximum tokens to predict (default: -1) - `--slots-endpoint-disable`: To disable slots state monitoring endpoint. Slots state may contain user data, prompts included. +- `--chat-template JINJA_TEMPLATE`: Set custom jinja chat template. This parameter accepts a string, not a file name (default: template taken from model's metadata). We only support [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) ## Build diff --git a/llama.h b/llama.h index 8ba20696f..84f196b3b 100644 --- a/llama.h +++ b/llama.h @@ -708,7 +708,7 @@ extern "C" { /// Apply chat template. Inspired by hf apply_chat_template() on python. /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model" - /// NOTE: This function only support some known jinja templates. It is not a jinja parser. + /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template /// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead. /// @param chat Pointer to a list of multiple llama_chat_message /// @param n_msg Number of llama_chat_message in this chat From 973053d8b0d04809836b3339a50f68d9c842de90 Mon Sep 17 00:00:00 2001 From: slaren Date: Thu, 22 Feb 2024 00:42:09 +0100 Subject: [PATCH 04/13] llama : fix loading models with shared tok_embd and output (#5651) ggml-ci --- llama.cpp | 12 ++++-------- 1 file changed, 4 insertions(+), 8 deletions(-) diff --git a/llama.cpp b/llama.cpp index d763cc80c..259f2a3a3 100644 --- a/llama.cpp +++ b/llama.cpp @@ -2791,13 +2791,7 @@ struct llama_model_loader { std::vector> read_buf; - for (int i = 0; i < gguf_get_n_tensors(ctx_gguf); i++) { - struct ggml_tensor * cur = ggml_get_tensor(ctx, gguf_get_tensor_name(ctx_gguf, i)); - if (!cur) { - // some tensors may be allocated in a different context - continue; - } - + for (struct ggml_tensor * cur = ggml_get_first_tensor(ctx); cur != NULL; cur = ggml_get_next_tensor(ctx, cur)) { if (progress_callback) { if (!progress_callback((float) size_done / size_data, progress_callback_user_data)) { return false; @@ -3722,7 +3716,7 @@ static bool llm_load_tensors( } // create one context per buffer type - size_t ctx_size = ggml_tensor_overhead()*ml.n_tensors; + size_t ctx_size = ggml_tensor_overhead()*(ml.n_tensors + 1); // +1 for models where tok_embd is duplicated as output std::map ctx_map; for (auto & it : buft_layer_count) { struct ggml_init_params params = { @@ -3860,6 +3854,7 @@ static bool llm_load_tensors( } else { model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // needs to be on GPU ml.n_created--; // artificial tensor + ml.size_data += ggml_nbytes(model.output); } } @@ -4396,6 +4391,7 @@ static bool llm_load_tensors( model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // same as tok_embd, duplicated to allow offloading ml.n_created--; // artificial tensor + ml.size_data += ggml_nbytes(model.output); const int64_t n_ff = hparams.n_ff; const int64_t n_embd_head_k = hparams.n_embd_head_k; From 4ef245a92a968ba0f18a5adfd41e51980ce4fdf5 Mon Sep 17 00:00:00 2001 From: Dat Quoc Nguyen <2412555+datquocnguyen@users.noreply.github.com> Date: Thu, 22 Feb 2024 18:15:13 +1000 Subject: [PATCH 05/13] mpt : add optional bias tensors (#5638) Update for MPT with optional bias parameters: to work with PhoGPT and SEA-LION models that were pre-trained with 'bias'. --- llama.cpp | 38 +++++++++++++++++++++++++++----------- 1 file changed, 27 insertions(+), 11 deletions(-) diff --git a/llama.cpp b/llama.cpp index 259f2a3a3..9cae8c761 100644 --- a/llama.cpp +++ b/llama.cpp @@ -4054,6 +4054,8 @@ static bool llm_load_tensors( // output { model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, false); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); } @@ -4063,14 +4065,23 @@ static bool llm_load_tensors( auto & layer = model.layers[i]; - layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, false); layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}); - layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, false); - layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); - layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}); - layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, false); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, false); + + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}); + layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, false); + + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, false); // AWQ ScaleActivation layer layer.ffn_act = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_ACT, "scales", i), {n_ff}, false); @@ -6171,7 +6182,7 @@ struct llm_build_context { attn_norm = llm_build_norm(ctx0, inpL, hparams, model.layers[il].attn_norm, - NULL, + model.layers[il].attn_norm_b, LLM_NORM, cb, il); cb(attn_norm, "attn_norm", il); @@ -6181,6 +6192,11 @@ struct llm_build_context { cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); cb(cur, "wqkv", il); + + if (model.layers[il].bqkv){ + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + } if (hparams.f_clamp_kqv > 0.0f) { cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); @@ -6198,7 +6214,7 @@ struct llm_build_context { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, - model.layers[il].wo, NULL, + model.layers[il].wo, model.layers[il].bo, Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -6211,13 +6227,13 @@ struct llm_build_context { { cur = llm_build_norm(ctx0, ffn_inp, hparams, model.layers[il].ffn_norm, - NULL, + model.layers[il].ffn_norm_b, LLM_NORM, cb, il); cb(cur, "ffn_norm", il); cur = llm_build_ffn(ctx0, cur, - model.layers[il].ffn_up, NULL, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, NULL, - model.layers[il].ffn_down, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, model.layers[il].ffn_act, LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); cb(cur, "ffn_out", il); @@ -6234,7 +6250,7 @@ struct llm_build_context { cur = llm_build_norm(ctx0, cur, hparams, model.output_norm, - NULL, + model.output_norm_b, LLM_NORM, cb, -1); cb(cur, "result_norm", -1); From c5688c6250430d2b8e0259efcf26c16dfa4c1f46 Mon Sep 17 00:00:00 2001 From: Alexey Parfenov Date: Thu, 22 Feb 2024 08:27:32 +0000 Subject: [PATCH 06/13] server : clarify some params in the docs (#5640) --- examples/server/README.md | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/examples/server/README.md b/examples/server/README.md index 4b24ee5dc..4b6cd8326 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -151,7 +151,7 @@ node index.js `temperature`: Adjust the randomness of the generated text (default: 0.8). - `dynatemp_range`: Dynamic temperature range (default: 0.0, 0.0 = disabled). + `dynatemp_range`: Dynamic temperature range. The final temperature will be in the range of `[temperature - dynatemp_range; temperature + dynatemp_range]` (default: 0.0, 0.0 = disabled). `dynatemp_exponent`: Dynamic temperature exponent (default: 1.0). @@ -209,7 +209,7 @@ node index.js `slot_id`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot (default: -1) - `cache_prompt`: Save the prompt and generation for avoid reprocess entire prompt if a part of this isn't change (default: false) + `cache_prompt`: Re-use previously cached prompt from the last request if possible. This may prevent re-caching the prompt from scratch. (default: false) `system_prompt`: Change the system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime) @@ -242,7 +242,7 @@ Notice that each `probs` is an array of length `n_probs`. - `content`: Completion result as a string (excluding `stopping_word` if any). In case of streaming mode, will contain the next token as a string. - `stop`: Boolean for use with `stream` to check whether the generation has stopped (Note: This is not related to stopping words array `stop` from input options) -- `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model` +- `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`. These options may differ from the original ones in some way (e.g. bad values filtered out, strings converted to tokens, etc.). - `model`: The path to the model loaded with `-m` - `prompt`: The provided `prompt` - `stopped_eos`: Indicating whether the completion has stopped because it encountered the EOS token From a46f50747b2028f7f9c9883b26bfba12bf92556e Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Thu, 22 Feb 2024 09:33:24 +0100 Subject: [PATCH 07/13] server : fallback to chatml, add AlphaMonarch chat template (#5628) * server: fallback to chatml * add new chat template * server: add AlphaMonarch to test chat template * server: only check model template if there is no custom tmpl * remove TODO --- examples/server/server.cpp | 15 +++++++++++++++ llama.cpp | 9 +++++++++ tests/test-chat-template.cpp | 23 +++++++++++++++-------- 3 files changed, 39 insertions(+), 8 deletions(-) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index c84719a0d..369121e88 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -400,6 +400,16 @@ struct llama_server_context return true; } + void validate_model_chat_template(server_params & sparams) { + llama_chat_message chat[] = {{"user", "test"}}; + std::vector buf(1); + int res = llama_chat_apply_template(model, nullptr, chat, 1, true, buf.data(), buf.size()); + if (res < 0) { + LOG_ERROR("The chat template comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses", {}); + sparams.chat_template = "<|im_start|>"; // llama_chat_apply_template only checks if <|im_start|> exist in the template + } + } + void initialize() { // create slots all_slots_are_idle = true; @@ -2752,6 +2762,11 @@ int main(int argc, char **argv) LOG_INFO("model loaded", {}); } + if (sparams.chat_template.empty()) { // custom chat template is not supplied + // check if the template comes with the model is supported by us + llama.validate_model_chat_template(sparams); + } + // Middleware for API key validation auto validate_api_key = [&sparams](const httplib::Request &req, httplib::Response &res) -> bool { // If API key is not set, skip validation diff --git a/llama.cpp b/llama.cpp index 9cae8c761..055b57e31 100644 --- a/llama.cpp +++ b/llama.cpp @@ -12773,6 +12773,15 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "<|assistant|>\n"; } + } else if (tmpl.find("bos_token + message['role']") != std::string::npos) { + // mlabonne/AlphaMonarch-7B template (the is included inside history) + for (auto message : chat) { + std::string bos = (message == chat.front()) ? "" : ""; // skip BOS for first message + ss << bos << message->role << "\n" << message->content << "\n"; + } + if (add_ass) { + ss << "assistant\n"; + } } else { // template not supported return -1; diff --git a/tests/test-chat-template.cpp b/tests/test-chat-template.cpp index 9830650d4..d02b39e14 100644 --- a/tests/test-chat-template.cpp +++ b/tests/test-chat-template.cpp @@ -27,12 +27,20 @@ int main(void) { "{%- for idx in range(0, messages|length) -%}\\n{%- if messages[idx]['role'] == 'user' -%}\\n{%- if idx > 1 -%}\\n{{- bos_token + '[INST] ' + messages[idx]['content'] + ' [/INST]' -}}\\n{%- else -%}\\n{{- messages[idx]['content'] + ' [/INST]' -}}\\n{%- endif -%}\\n{% elif messages[idx]['role'] == 'system' %}\\n{{- '[INST] <>\\\\n' + messages[idx]['content'] + '\\\\n<>\\\\n\\\\n' -}}\\n{%- elif messages[idx]['role'] == 'assistant' -%}\\n{{- ' ' + messages[idx]['content'] + ' ' + eos_token -}}\\n{% endif %}\\n{% endfor %}", // bofenghuang/vigogne-2-70b-chat "{{ bos_token }}{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% elif true == true and not '<>' in messages[0]['content'] %}{% set loop_messages = messages %}{% set system_message = 'Vous êtes Vigogne, un assistant IA créé par Zaion Lab. Vous suivez extrêmement bien les instructions. Aidez autant que vous le pouvez.' %}{% else %}{% set loop_messages = messages %}{% set system_message = false %}{% endif %}{% for message in loop_messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if loop.index0 == 0 and system_message != false %}{% set content = '<>\\\\n' + system_message + '\\\\n<>\\\\n\\\\n' + message['content'] %}{% else %}{% set content = message['content'] %}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + content.strip() + ' [/INST]' }}{% elif message['role'] == 'system' %}{{ '<>\\\\n' + content.strip() + '\\\\n<>\\\\n\\\\n' }}{% elif message['role'] == 'assistant' %}{{ ' ' + content.strip() + ' ' + eos_token }}{% endif %}{% endfor %}", + // mlabonne/AlphaMonarch-7B + "{% for message in messages %}{{bos_token + message['role'] + '\\n' + message['content'] + eos_token + '\\n'}}{% endfor %}{% if add_generation_prompt %}{{ bos_token + 'assistant\\n' }}{% endif %}", }; - std::vector expected_substr = { - "<|im_start|>assistant\n I am an assistant <|im_end|>\n<|im_start|>user\nAnother question<|im_end|>\n<|im_start|>assistant", - "[/INST]Hi there[INST] Who are you [/INST] I am an assistant [INST] Another question [/INST]", - "[INST] Who are you [/INST] I am an assistant [INST] Another question [/INST]", - "[/INST] Hi there [INST] Who are you [/INST] I am an assistant [INST] Another question [/INST]", + std::vector expected_output = { + // teknium/OpenHermes-2.5-Mistral-7B + "<|im_start|>system\nYou are a helpful assistant<|im_end|>\n<|im_start|>user\nHello<|im_end|>\n<|im_start|>assistant\nHi there<|im_end|>\n<|im_start|>user\nWho are you<|im_end|>\n<|im_start|>assistant\n I am an assistant <|im_end|>\n<|im_start|>user\nAnother question<|im_end|>\n<|im_start|>assistant\n", + // mistralai/Mistral-7B-Instruct-v0.2 + "[INST] You are a helpful assistant\nHello [/INST]Hi there[INST] Who are you [/INST] I am an assistant [INST] Another question [/INST]", + // TheBloke/FusionNet_34Bx2_MoE-AWQ + "[INST] <>\nYou are a helpful assistant\n<>\n\nHello [/INST] Hi there [INST] Who are you [/INST] I am an assistant [INST] Another question [/INST]", + // bofenghuang/vigogne-2-70b-chat + "[INST] <>\nYou are a helpful assistant\n<>\n\nHello [/INST] Hi there [INST] Who are you [/INST] I am an assistant [INST] Another question [/INST]", + // mlabonne/AlphaMonarch-7B + "system\nYou are a helpful assistant\nuser\nHello\nassistant\nHi there\nuser\nWho are you\nassistant\n I am an assistant \nuser\nAnother question\nassistant\n", }; std::vector formatted_chat(1024); int32_t res; @@ -43,7 +51,7 @@ int main(void) { for (size_t i = 0; i < templates.size(); i++) { std::string custom_template = templates[i]; - std::string substr = expected_substr[i]; + std::string expected = expected_output[i]; formatted_chat.resize(1024); res = llama_chat_apply_template( nullptr, @@ -57,8 +65,7 @@ int main(void) { formatted_chat.resize(res); std::string output(formatted_chat.data(), formatted_chat.size()); std::cout << output << "\n-------------------------\n"; - // expect the "formatted_chat" to contain pre-defined strings - assert(output.find(substr) != std::string::npos); + assert(output == expected); } return 0; } From 56d03d92be57f5880b9ed94542d87bb6effae31f Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 22 Feb 2024 10:35:54 +0200 Subject: [PATCH 08/13] readme : update hot topics --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index c1624b9f9..3bc512af0 100644 --- a/README.md +++ b/README.md @@ -10,6 +10,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) ### Hot topics +- Support for chat templates: [Wiki (contributions welcome)](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) - Support for Gemma models: https://github.com/ggerganov/llama.cpp/pull/5631 - Non-linear quantization IQ4_NL: https://github.com/ggerganov/llama.cpp/pull/5590 - Looking for contributions to improve and maintain the `server` example: https://github.com/ggerganov/llama.cpp/issues/4216 From 3a03541cedea474fa9d41214484cc3fbcf468a9e Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 22 Feb 2024 13:54:03 +0200 Subject: [PATCH 09/13] minor : fix trailing whitespace (#5638) --- llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama.cpp b/llama.cpp index 055b57e31..6ab5e1bf4 100644 --- a/llama.cpp +++ b/llama.cpp @@ -6192,7 +6192,7 @@ struct llm_build_context { cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); cb(cur, "wqkv", il); - + if (model.layers[il].bqkv){ cur = ggml_add(ctx0, cur, model.layers[il].bqkv); cb(cur, "bqkv", il); From 4cb4d8b22d4fda971621a68c570ce84d66897c37 Mon Sep 17 00:00:00 2001 From: Someone Date: Thu, 22 Feb 2024 16:32:09 +0000 Subject: [PATCH 10/13] workflows: nix: hardcode cachix ids, build unconditionally (#5663) GitHub does not expose environment and repository variables to PRs coming from forks implies that we've been disabling the Nix CI actions for most PRs. The `if:` also didn't make much sense, because we can always pull from cachix, and there's no point (albeit no risk either) in pushing cache for the untrusted code. --- .github/workflows/nix-ci-aarch64.yml | 7 +++---- .github/workflows/nix-ci.yml | 11 +++++------ 2 files changed, 8 insertions(+), 10 deletions(-) diff --git a/.github/workflows/nix-ci-aarch64.yml b/.github/workflows/nix-ci-aarch64.yml index 0c6cf5f09..8d0a3fd7f 100644 --- a/.github/workflows/nix-ci-aarch64.yml +++ b/.github/workflows/nix-ci-aarch64.yml @@ -19,7 +19,6 @@ on: jobs: nix-build-aarch64: - if: ${{ vars.CACHIX_NAME != '' }} runs-on: ubuntu-latest steps: - name: Checkout repository @@ -37,8 +36,8 @@ jobs: extra-conf: | extra-platforms = aarch64-linux extra-system-features = nixos-test kvm - extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org - extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E= + extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org + extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E= - uses: DeterminateSystems/magic-nix-cache-action@v2 with: upstream-cache: https://${{ matrix.cachixName }}.cachix.org @@ -46,7 +45,7 @@ jobs: uses: cachix/cachix-action@v13 with: authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}' - name: ${{ vars.CACHIX_NAME }} + name: llama-cpp - name: Show all output paths run: > nix run github:nix-community/nix-eval-jobs diff --git a/.github/workflows/nix-ci.yml b/.github/workflows/nix-ci.yml index d19c7a576..01c5a9d5a 100644 --- a/.github/workflows/nix-ci.yml +++ b/.github/workflows/nix-ci.yml @@ -23,8 +23,8 @@ jobs: with: github-token: ${{ secrets.GITHUB_TOKEN }} extra-conf: | - extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org - extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E= + extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org + extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E= - uses: DeterminateSystems/magic-nix-cache-action@v2 with: upstream-cache: https://${{ matrix.cachixName }}.cachix.org @@ -37,7 +37,6 @@ jobs: --flake ".#packages.$(nix eval --raw --impure --expr builtins.currentSystem)" nix-build: - if: ${{ vars.CACHIX_NAME != '' }} strategy: fail-fast: false matrix: @@ -51,8 +50,8 @@ jobs: with: github-token: ${{ secrets.GITHUB_TOKEN }} extra-conf: | - extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org - extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E= + extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org + extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E= - uses: DeterminateSystems/magic-nix-cache-action@v2 with: upstream-cache: https://${{ matrix.cachixName }}.cachix.org @@ -60,7 +59,7 @@ jobs: uses: cachix/cachix-action@v13 with: authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}' - name: ${{ vars.CACHIX_NAME }} + name: llama-cpp - name: Build run: > nix run github:Mic92/nix-fast-build From 373ee3fbbabc4c1508eed4f5c3795b23a20939a3 Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Thu, 22 Feb 2024 19:10:21 +0100 Subject: [PATCH 11/13] Add Gemma chat template (#5665) * add gemma chat template * gemma: only apply system_prompt on non-model message --- llama.cpp | 22 ++++++++++++++++++++++ tests/test-chat-template.cpp | 4 ++++ 2 files changed, 26 insertions(+) diff --git a/llama.cpp b/llama.cpp index 6ab5e1bf4..40dda265c 100644 --- a/llama.cpp +++ b/llama.cpp @@ -12782,6 +12782,28 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "assistant\n"; } + } else if (tmpl.find("") != std::string::npos) { + // google/gemma-7b-it + std::string system_prompt = ""; + for (auto message : chat) { + std::string role(message->role); + if (role == "system") { + // there is no system message for gemma, but we will merge it with user prompt, so nothing is broken + system_prompt = trim(message->content); + continue; + } + // in gemma, "assistant" is "model" + role = role == "assistant" ? "model" : message->role; + ss << "" << role << "\n"; + if (!system_prompt.empty() && role != "model") { + ss << system_prompt << "\n\n"; + system_prompt = ""; + } + ss << trim(message->content) << "\n"; + } + if (add_ass) { + ss << "model\n"; + } } else { // template not supported return -1; diff --git a/tests/test-chat-template.cpp b/tests/test-chat-template.cpp index d02b39e14..fa2eb577b 100644 --- a/tests/test-chat-template.cpp +++ b/tests/test-chat-template.cpp @@ -29,6 +29,8 @@ int main(void) { "{{ bos_token }}{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% elif true == true and not '<>' in messages[0]['content'] %}{% set loop_messages = messages %}{% set system_message = 'Vous êtes Vigogne, un assistant IA créé par Zaion Lab. Vous suivez extrêmement bien les instructions. Aidez autant que vous le pouvez.' %}{% else %}{% set loop_messages = messages %}{% set system_message = false %}{% endif %}{% for message in loop_messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if loop.index0 == 0 and system_message != false %}{% set content = '<>\\\\n' + system_message + '\\\\n<>\\\\n\\\\n' + message['content'] %}{% else %}{% set content = message['content'] %}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + content.strip() + ' [/INST]' }}{% elif message['role'] == 'system' %}{{ '<>\\\\n' + content.strip() + '\\\\n<>\\\\n\\\\n' }}{% elif message['role'] == 'assistant' %}{{ ' ' + content.strip() + ' ' + eos_token }}{% endif %}{% endfor %}", // mlabonne/AlphaMonarch-7B "{% for message in messages %}{{bos_token + message['role'] + '\\n' + message['content'] + eos_token + '\\n'}}{% endfor %}{% if add_generation_prompt %}{{ bos_token + 'assistant\\n' }}{% endif %}", + // google/gemma-7b-it + "{% if messages[0]['role'] == 'system' %}{{ raise_exception('System role not supported') }}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = message['role'] %}{% endif %}{{ '' + role + '\\n' + message['content'] | trim + '\\n' }}{% endfor %}{% if add_generation_prompt %}{{'model\\n'}}{% endif %}", }; std::vector expected_output = { // teknium/OpenHermes-2.5-Mistral-7B @@ -41,6 +43,8 @@ int main(void) { "[INST] <>\nYou are a helpful assistant\n<>\n\nHello [/INST] Hi there [INST] Who are you [/INST] I am an assistant [INST] Another question [/INST]", // mlabonne/AlphaMonarch-7B "system\nYou are a helpful assistant\nuser\nHello\nassistant\nHi there\nuser\nWho are you\nassistant\n I am an assistant \nuser\nAnother question\nassistant\n", + // google/gemma-7b-it + "user\nYou are a helpful assistant\n\nHello\nmodel\nHi there\nuser\nWho are you\nmodel\nI am an assistant\nuser\nAnother question\nmodel\n", }; std::vector formatted_chat(1024); int32_t res; From 5a9e2f60ba3d8362ba17c77ac3092906d49b813f Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 22 Feb 2024 20:13:25 +0200 Subject: [PATCH 12/13] py : minor fixes (#5668) --- convert-hf-to-gguf.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 9771fccf9..8630bbf29 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -655,6 +655,8 @@ class OrionModel(Model): self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"]) self.gguf_writer.add_head_count(head_count) self.gguf_writer.add_head_count_kv(head_count_kv) + # note: config provides rms norm but it is actually layer norm + # ref: https://huggingface.co/OrionStarAI/Orion-14B-Chat/blob/276a17221ce42beb45f66fac657a41540e71f4f5/modeling_orion.py#L570-L571 self.gguf_writer.add_layer_norm_eps(self.hparams["rms_norm_eps"]) def write_tensors(self): @@ -1031,7 +1033,6 @@ class PersimmonModel(Model): self.gguf_writer.add_head_count_kv(head_count_kv) self.gguf_writer.add_rope_freq_base(self.hparams["rope_theta"]) self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"]) - self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"]) def set_vocab(self): self._set_vocab_sentencepiece() From 201294ae177b308fb3a99dc504dd6d27e8afa907 Mon Sep 17 00:00:00 2001 From: Someone Date: Thu, 22 Feb 2024 19:44:10 +0000 Subject: [PATCH 13/13] nix: init singularity and docker images (#5056) Exposes a few attributes demonstrating how to build [singularity](https://docs.sylabs.io/guides/latest/user-guide/)/[apptainer](https://apptainer.org/) and Docker images re-using llama.cpp's Nix expression. Built locally on `x86_64-linux` with `nix build github:someoneserge/llama.cpp/feat/nix/images#llamaPackages.{docker,docker-min,sif,llama-cpp}` and it's fast and effective. --- .devops/nix/docker.nix | 37 +++++++++++++++++++++++++++++++++++++ .devops/nix/scope.nix | 3 +++ .devops/nix/sif.nix | 27 +++++++++++++++++++++++++++ 3 files changed, 67 insertions(+) create mode 100644 .devops/nix/docker.nix create mode 100644 .devops/nix/sif.nix diff --git a/.devops/nix/docker.nix b/.devops/nix/docker.nix new file mode 100644 index 000000000..d607b4575 --- /dev/null +++ b/.devops/nix/docker.nix @@ -0,0 +1,37 @@ +{ + lib, + dockerTools, + buildEnv, + llama-cpp, + interactive ? true, + coreutils, +}: + +# A tar that can be fed into `docker load`: +# +# $ nix build .#llamaPackages.docker +# $ docker load < result + +# For details and variations cf. +# - https://nixos.org/manual/nixpkgs/unstable/#ssec-pkgs-dockerTools-buildLayeredImage +# - https://discourse.nixos.org/t/a-faster-dockertools-buildimage-prototype/16922 +# - https://nixery.dev/ + +# Approximate (compressed) sizes, at the time of writing, are: +# +# .#llamaPackages.docker: 125M; +# .#llamaPackagesCuda.docker: 537M; +# .#legacyPackages.aarch64-linux.llamaPackagesXavier.docker: 415M. + +dockerTools.buildLayeredImage { + name = llama-cpp.pname; + tag = "latest"; + + contents = + [ llama-cpp ] + ++ lib.optionals interactive [ + coreutils + dockerTools.binSh + dockerTools.caCertificates + ]; +} diff --git a/.devops/nix/scope.nix b/.devops/nix/scope.nix index d295995a4..78530c9e8 100644 --- a/.devops/nix/scope.nix +++ b/.devops/nix/scope.nix @@ -12,5 +12,8 @@ lib.makeScope newScope ( self: { inherit llamaVersion; llama-cpp = self.callPackage ./package.nix { }; + docker = self.callPackage ./docker.nix { }; + docker-min = self.callPackage ./docker.nix { interactive = false; }; + sif = self.callPackage ./sif.nix { }; } ) diff --git a/.devops/nix/sif.nix b/.devops/nix/sif.nix new file mode 100644 index 000000000..7535ca0f3 --- /dev/null +++ b/.devops/nix/sif.nix @@ -0,0 +1,27 @@ +{ + lib, + singularity-tools, + llama-cpp, + bashInteractive, + interactive ? false, +}: + +let + optionalInt = cond: x: if cond then x else 0; +in +singularity-tools.buildImage rec { + inherit (llama-cpp) name; + contents = [ llama-cpp ] ++ lib.optionals interactive [ bashInteractive ]; + + # These are excessive (but safe) for most variants. Building singularity + # images requires superuser privileges, so we build them inside a VM in a + # writable image of pre-determined size. + # + # ROCm is currently affected by https://github.com/NixOS/nixpkgs/issues/276846 + # + # Expected image sizes: + # - cpu/blas: 150M, + # - cuda, all gencodes: 560M, + diskSize = 4096 + optionalInt llama-cpp.useRocm 16384; + memSize = diskSize; +}