From 46b3ccaa6ff60888f38f122d9cd6670f0003517d Mon Sep 17 00:00:00 2001 From: willhe Date: Wed, 13 Mar 2024 11:25:59 +0000 Subject: [PATCH] 1. Convert xverse models to gguf; 2. Add LLM_ARCH_XVERSE inference in llama.cpp; 3. Add xverse item in Supported models in README.md; --- README.md | 1 + convert-hf-to-gguf.py | 33 ++++---- llama.cpp | 173 +++++++++++++++++++++++++++++++++++++++++- 3 files changed, 192 insertions(+), 15 deletions(-) diff --git a/README.md b/README.md index 54bf84bec..54805bb7a 100644 --- a/README.md +++ b/README.md @@ -110,6 +110,7 @@ Typically finetunes of the base models below are supported as well. - [x] [CodeShell](https://github.com/WisdomShell/codeshell) - [x] [Gemma](https://ai.google.dev/gemma) - [x] [Mamba](https://github.com/state-spaces/mamba) +- [x] [Xverse](https://huggingface.co/models?search=xverse) **Multimodal models:** diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 5e065d39b..955c09eb0 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -772,6 +772,7 @@ class XverseModel(Model): assert (self.dir_model / "tokenizer.json").is_file() dir_model = self.dir_model hparams = self.hparams + tokens: list[bytearray] = [] toktypes: list[int] = [] @@ -783,26 +784,30 @@ class XverseModel(Model): reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()} added_vocab = tokenizer.get_added_vocab() - for i in range(vocab_size): - if i not in reverse_vocab: - pad_token = f"[PAD{i}]".encode('utf-8') - tokens.append(bytearray(pad_token)) - toktypes.append(gguf.TokenType.USER_DEFINED) - elif reverse_vocab[i] in added_vocab: - tokens.append(reverse_vocab[i]) - if tokenizer.added_tokens_decoder[i].special: - toktypes.append(gguf.TokenType.CONTROL) + for token_id in range(vocab_size): + token_text = reverse_vocab[token_id].encode('utf-8') + # replace "\x00" to string with length > 0 + if token_text == b"\x00": + toktype = gguf.TokenType.BYTE # special + token_text = f"<{token_text}>".encode('utf-8') + elif re.fullmatch(br"<0x[0-9A-Fa-f]{2}>", token_text): + toktype = gguf.TokenType.BYTE # special + elif reverse_vocab[token_id] in added_vocab: + if tokenizer.added_tokens_decoder[token_id].special: + toktype = gguf.TokenType.CONTROL else: - toktypes.append(gguf.TokenType.USER_DEFINED) + toktype = gguf.TokenType.USER_DEFINED else: - tokens.append(reverse_vocab[i]) - toktypes.append(gguf.TokenType.NORMAL) + toktype = gguf.TokenType.NORMAL - self.gguf_writer.add_tokenizer_model("xverse") + tokens.append(token_text) + toktypes.append(toktype) + + self.gguf_writer.add_tokenizer_model("llama") self.gguf_writer.add_token_list(tokens) self.gguf_writer.add_token_types(toktypes) - special_vocab = gguf.SpecialVocab(dir_model, load_merges=True) + special_vocab = gguf.SpecialVocab(dir_model, n_vocab=len(tokens)) special_vocab.add_to_gguf(self.gguf_writer) def set_gguf_parameters(self): diff --git a/llama.cpp b/llama.cpp index ad7b7b7d4..39349a2a6 100644 --- a/llama.cpp +++ b/llama.cpp @@ -214,6 +214,7 @@ enum llm_arch { LLM_ARCH_GEMMA, LLM_ARCH_STARCODER2, LLM_ARCH_MAMBA, + LLM_ARCH_XVERSE, LLM_ARCH_UNKNOWN, }; @@ -243,6 +244,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_GEMMA, "gemma" }, { LLM_ARCH_STARCODER2, "starcoder2" }, { LLM_ARCH_MAMBA, "mamba" }, + { LLM_ARCH_XVERSE, "xverse" }, { LLM_ARCH_UNKNOWN, "(unknown)" }, }; @@ -836,6 +838,25 @@ static const std::map> LLM_TENSOR_NA { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" }, }, }, + { + LLM_ARCH_XVERSE, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, { LLM_ARCH_UNKNOWN, { @@ -3628,6 +3649,16 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_XVERSE: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_7B; break; + case 40: model.type = e_model::MODEL_13B; break; + case 80: model.type = e_model::MODEL_65B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; default: (void)0; } @@ -4106,6 +4137,7 @@ static bool llm_load_tensors( LLAMA_LOG_INFO("%s: ggml ctx size = %7.2f MiB\n", __func__, model.ctxs.size()*ctx_size/1024.0/1024.0); + bool init_mapping_prefetch = true; // create tensors for the weights { const int64_t n_embd = hparams.n_embd; @@ -4910,6 +4942,35 @@ static bool llm_load_tensors( layer.ssm_out = ml.create_tensor(ctx_split, tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}); } } break; + case LLM_ARCH_XVERSE: + { + init_mapping_prefetch = false; + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + { + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); + } + + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); + layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); + layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + + layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); + 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}); + } + } break; default: throw std::runtime_error("unknown architecture"); } @@ -4917,7 +4978,7 @@ static bool llm_load_tensors( ml.done_getting_tensors(); - ml.init_mapping(true, use_mlock ? &model.mlock_mmap : nullptr); + ml.init_mapping(init_mapping_prefetch, use_mlock ? &model.mlock_mmap : nullptr); // create the backend buffers std::vector> ctx_bufs; @@ -5910,6 +5971,111 @@ struct llm_build_context { return gf; } + struct ggml_cgraph * build_xverse() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); + cb(inpL, "inp_embd", -1); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); + cb(inp_pos, "inp_pos", -1); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); + cb(KQ_mask, "KQ_mask", -1); + + // positions of the tokens in the KV cache + struct ggml_tensor * KQ_pos = ggml_view_1d(ctx0, lctx.inp_KQ_pos, n_kv, 0); + cb(KQ_pos, "KQ_pos", -1); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + model.layers[il].wo, NULL, + 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); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, + model.layers[il].ffn_gate, NULL, + model.layers[il].ffn_down, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = llm_build_norm(ctx0, cur, hparams, model.output_norm, NULL, LLM_NORM_RMS, cb, -1); + cb(cur, "result_norm", -1); + + // lm_head + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } + struct ggml_cgraph * build_falcon() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); @@ -8473,6 +8639,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_mamba(); } break; + case LLM_ARCH_XVERSE: + { + result = llm.build_xverse(); + } break; default: GGML_ASSERT(false); } @@ -13053,6 +13223,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { case LLM_ARCH_ORION: case LLM_ARCH_INTERNLM2: case LLM_ARCH_MINICPM: + case LLM_ARCH_XVERSE: return LLAMA_ROPE_TYPE_NORM; // the pairs of head values are offset by n_rot/2