概述
本文在qemu-riscv64平台上，利用向量扩展指令加速运行基于llama.cpp构建的大模型。
参考博客链接：
Accelerating llama.cpp with RISC-V Vector Extension
基于RVV的llama.cpp在Banana Pi F3 RISCV开发板上的演示
llama.cpp工程
Llama.cpp是一个基于C++编写的高性能大模型推理框架，旨在提供快速、稳定且易于使用的计算工具，Llama.cpp支持多种计算模式，包括向量计算、矩阵运算、图算法等，可广泛应用于机器学习、图像处理、数据分析等领域。
目录结构
src是构建模型架构的基础库文件夹
examples是部分案例模型的源文件
ggml是计算操作的库文件
源码分析
可如下参考链接（注意：函数名有所变动）：
CodeLeaner@微信公众号：llama.cpp源码解析
llama模型
// llama @examples/main/main.cpp
main()
gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_MAIN, print_usage) // 解析传递进来的模型参数
llama_init_from_gpt_params()
llama_load_model_from_file(params.model.c_str(), mparams); // 加载model参数
llama_new_context_with_model(model, cparams); //
ggml_backend_cpu_init();
*cpu_backend // 定义指针指向 cpu_backend_i
llama_tokenize(ctx, prompt, true, true) // 将prompt tokenize
while // 循环产生token
llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0)) // 生成token函数
llama_token_to_piece(ctx, id, params.special)
gpt_perf_print(ctx, smpl); // 打印性能结果
llama.cpp模型库文件
// decode 函数体 @src/llama.cpp
int32_t llama_decode( struct llama_context * ctx, struct llama_batch batch)
llama_decode_internal(*ctx, batch)
while (lctx.sbatch.n_tokens > 0)
llama_build_graph(lctx, ubatch, false);
llama_graph_compute(lctx, gf, n_threads, threadpool);
ggml_backend_cpu_set_n_threads(lctx.backend_cpu, n_threads);
ggml_backend_cpu_set_threadpool(lctx.backend_cpu, threadpool);
ggml_backend_cpu_set_abort_callback(lctx.backend_cpu, lctx.abort_callback, lctx.abort_callback_data);
ggml_backend_sched_graph_compute_async(lctx.sched, gf);
ggml计算库函数
// backend计算图执行函数 ggml/src/ggml-backend.c
enum ggml_status ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph)
ggml_backend_sched_alloc_graph(sched, graph)
ggml_backend_sched_split_graph(sched, graph); // 该函数划分graph，并调用cpu_backend进行计算，继而调用ggml_graph_compute计算
// ggml计算图中各类计算选通的主体函数 ggml/src/ggml.c
enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan)
ggml_graph_compute_thread(&threadpool->workers[0])
ggml_compute_forward(¶ms, node);
ggml_compute_forward_dup(params, tensor);
ggml_compute_forward_add1(params, tensor);
ggml_compute_forward_repeat(params, tensor);
ggml_compute_forward_mul_mat(params, tensor);
ggml_compute_forward_soft_max(params, tensor);
ggml_compute_forward_rms_norm(params, tensor);
结构体
// cpu执行数据流结构体，将函数作为结构体成员。
cpu_backend_i // @ggml/src/ggml-backend.c
ggml_backend_cpu_graph_plan_create()
ggml_graph_plan()
ggml_backend_cpu_graph_plan_compute()
ggml_graph_compute()
ggml_backend_cpu_graph_compute()
ggml_graph_plan()
ggml_graph_compute()
llama.cpp中量化方式（Qn_0、Qn_1等）含义
sgsprog@hackmd.io: Linux 核心專題: llama.cpp 效能分析
rvv移植代码分析
Github相关链接：
Llama.cpp中利用GGML中对RVV的支持1
Llama.cpp中利用GGML中对RVV的支持2
Tameem-10xE@llama.cpp Github：Added RISC-V Vector Intrinsics Support
起初移植代码在ggml.c中后续迁移至ggml-quants.c文件中。
修改函数包括12个：
quantize_row_q8_0
quantize_row_q8_1
ggml_vec_dot_q4_0_q8_0
ggml_vec_dot_q4_1_q8_1
ggml_vec_dot_q5_0_q8_0
ggml_vec_dot_q5_1_q8_1
ggml_vec_dot_q8_0_q8_0
ggml_vec_dot_q2_K_q8_K
ggml_vec_dot_q3_K_q8_K
ggml_vec_dot_q4_K_q8_K
ggml_vec_dot_q5_K_q8_K
ggml_vec_dot_q6_K_q8_K
这些函数作为不同量化模式的成员函数:
static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
[GGML_TYPE_Q8_0] = {
.type_name = "q8_0",
.blck_size = QK8_0,
.type_size = sizeof(block_q8_0),
.is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_q8_0,
.from_float = quantize_row_q8_0,
.from_float_ref = (ggml_from_float_t) quantize_row_q8_0_ref,
.from_float_to_mat = quantize_mat_q8_0,
.vec_dot = ggml_vec_dot_q8_0_q8_0,
.vec_dot_type = GGML_TYPE_Q8_0,
......
}
在ggml不同算子计算函数中被调用：
static void ggml_compute_forward_mul_mat_one_chunk(...)
ggml_vec_dot_t const vec_dot = type_traits[type].vec_dot;
for (int64_t iir1 = ir1_start; iir1 < ir1_end; iir1 += blck_1) {
for (int64_t iir0 = ir0_start; iir0 < ir0_end; iir0 += blck_0) {
for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir1_end; ir1 += num_rows_per_vec_dot) {
...
for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir0_end; ir0 += num_rows_per_vec_dot) {
vec_dot(ne00, &tmp[ir0 - iir0], (num_rows_per_vec_dot > 1 ? 16 : 0), src0_row + ir0 * nb01, (num_rows_per_vec_dot > 1 ? nb01 : 0), src1_col, (num_rows_per_vec_dot > 1 ? src1_col_stride : 0), num_rows_per_vec_dot);
}
...
}
}
}
2024/10/02: 工具准备OK，但qemu运行时被killed
工具版本
Qemu：
Gcc版本：
Github Release
llama.cpp：
llama.cpp Github 10月2号pull
llama-7b模型版本：
Huggingface gguf文件
编译
llama.cpp编译
cd llama.cpp
make RISCV_CROSS_COMPILE=1
运行命令
qemu-riscv64 -L /home/kevin/data/projects/tools/riscv64_linux_gcc/sysroot -cpu rv64,v=true,vlen=256,elen=64,vext_spec=v1.0 ./llama-server -m /home/kevin/data/projects/kg_proj/rvv_transformer/codellama-7b.Q4_K_M.gguf -p “Anything” -n 9
问题
命令运行现象
可能原因
运行内存可能太小
2024/10/03: 使用10xE团队的最新版，解决tokenizer的问题，但还是被killed
最新版Github链接
Tameem-10xE/llama.cpp Github
问题：运行7B模型被killed
运行现象
可能原因
有可能跟qemu运行的swapfile有关
可以团队成员提的一个issue：
Github Issue： qemu-riscv64 unexpectedly reached EOF error
解决办法
先尝试换一个更小的模型试试，不行就解决swapfile的问题
运行3B规模的model的现象：failed to allocate buffer of size
kevin@BRICKHOUSE01:~/data/projects/kg_proj/rvv_transformer/llama.cpp$ qemu-riscv64 -L /home/kevin/data/projects/tools/riscv64_linux_gcc/sysroot -cpu rv64,v=true,vlen=256,elen=64,vext_spec=v1.0 ./llama-cli -m /home/kevin/data/projects/kg_proj/rvv_transformer/Llama-3.2-3B-Instruct-IQ3_M.gguf -p "Anything" -n 9
Log start
main: build = 3733 (e5701063)
main: built with riscv64-unknown-linux-gnu-gcc () 13.2.0 for riscv64-unknown-linux-gnu
llama_model_loader: loaded meta data with 35 key-value pairs and 255 tensors from /home/kevin/data/projects/kg_proj/rvv_transformer/Llama-3.2-3B-Instruct-IQ3_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv 0: general.architecture str = llama
llama_model_loader: - kv 1: general.type str = model
llama_model_loader: - kv 2: general.name str = Llama 3.2 3B Instruct
llama_model_loader: - kv 3: general.finetune str = Instruct
llama_model_loader: - kv 4: general.basename str = Llama-3.2
llama_model_loader: - kv 5: general.size_label str = 3B
llama_model_loader: - kv 6: general.license str = llama3.2
llama_model_loader: - kv 7: general.tags arr[str,6] = ["facebook", "meta", "pytorch", "llam...
llama_model_loader: - kv 8: general.languages arr[str,8] = ["en", "de", "fr", "it", "pt", "hi", ...
llama_model_loader: - kv 9: llama.block_count u32 = 28
llama_model_loader: - kv 10: llama.context_length u32 = 131072
llama_model_loader: - kv 11: llama.embedding_length u32 = 3072
llama_model_loader: - kv 12: llama.feed_forward_length u32 = 8192
llama_model_loader: - kv 13: llama.attention.head_count u32 = 24
llama_model_loader: - kv 14: llama.attention.head_count_kv u32 = 8
llama_model_loader: - kv 15: llama.rope.freq_base f32 = 500000.000000
llama_model_loader: - kv 16: llama.attention.layer_norm_rms_epsilon f32 = 0.000010
llama_model_loader: - kv 17: llama.attention.key_length u32 = 128
llama_model_loader: - kv 18: llama.attention.value_length u32 = 128
llama_model_loader: - kv 19: general.file_type u32 = 27
llama_model_loader: - kv 20: llama.vocab_size u32 = 128256
llama_model_loader: - kv 21: llama.rope.dimension_count u32 = 128
llama_model_loader: - kv 22: tokenizer.ggml.model str = gpt2
llama_model_loader: - kv 23: tokenizer.ggml.pre str = llama-bpe
llama_model_loader: - kv 24: tokenizer.ggml.tokens arr[str,128256] = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv 25: tokenizer.ggml.token_type arr[i32,128256] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 26: tokenizer.ggml.merges arr[str,280147] = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
llama_model_loader: - kv 27: tokenizer.ggml.bos_token_id u32 = 128000
llama_model_loader: - kv 28: tokenizer.ggml.eos_token_id u32 = 128009
llama_model_loader: - kv 29: tokenizer.chat_template str = {{- bos_token }}\n{%- if custom_tools ...
llama_model_loader: - kv 30: general.quantization_version u32 = 2
llama_model_loader: - kv 31: quantize.imatrix.file str = /models_out/Llama-3.2-3B-Instruct-GGU...
llama_model_loader: - kv 32: quantize.imatrix.dataset str = /training_dir/calibration_datav3.txt
llama_model_loader: - kv 33: quantize.imatrix.entries_count i32 = 196
llama_model_loader: - kv 34: quantize.imatrix.chunks_count i32 = 125
llama_model_loader: - type f32: 58 tensors
llama_model_loader: - type q4_K: 59 tensors
llama_model_loader: - type q6_K: 1 tensors
llama_model_loader: - type iq3_s: 137 tensors
llm_load_vocab: special tokens cache size = 256
llm_load_vocab: token to piece cache size = 0.7999 MB
llm_load_print_meta: format = GGUF V3 (latest)
llm_load_print_meta: arch = llama
llm_load_print_meta: vocab type = BPE
llm_load_print_meta: n_vocab = 128256
llm_load_print_meta: n_merges = 280147
llm_load_print_meta: vocab_only = 0
llm_load_print_meta: n_ctx_train = 131072
llm_load_print_meta: n_embd = 3072
llm_load_print_meta: n_layer = 28
llm_load_print_meta: n_head = 24
llm_load_print_meta: n_head_kv = 8
llm_load_print_meta: n_rot = 128
llm_load_print_meta: n_swa = 0
llm_load_print_meta: n_embd_head_k = 128
llm_load_print_meta: n_embd_head_v = 128
llm_load_print_meta: n_gqa = 3
llm_load_print_meta: n_embd_k_gqa = 1024
llm_load_print_meta: n_embd_v_gqa = 1024
llm_load_print_meta: f_norm_eps = 0.0e+00
llm_load_print_meta: f_norm_rms_eps = 1.0e-05
llm_load_print_meta: f_clamp_kqv = 0.0e+00
llm_load_print_meta: f_max_alibi_bias = 0.0e+00
llm_load_print_meta: f_logit_scale = 0.0e+00
llm_load_print_meta: n_ff = 8192
llm_load_print_meta: n_expert = 0
llm_load_print_meta: n_expert_used = 0
llm_load_print_meta: causal attn = 1
llm_load_print_meta: pooling type = 0
llm_load_print_meta: rope type = 0
llm_load_print_meta: rope scaling = linear
llm_load_print_meta: freq_base_train = 500000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn = 131072
llm_load_print_meta: rope_finetuned = unknown
llm_load_print_meta: ssm_d_conv = 0
llm_load_print_meta: ssm_d_inner = 0
llm_load_print_meta: ssm_d_state = 0
llm_load_print_meta: ssm_dt_rank = 0
llm_load_print_meta: ssm_dt_b_c_rms = 0
llm_load_print_meta: model type = ?B
llm_load_print_meta: model ftype = IQ3_S mix - 3.66 bpw
llm_load_print_meta: model params = 3.21 B
llm_load_print_meta: model size = 1.48 GiB (3.96 BPW)
llm_load_print_meta: general.name = Llama 3.2 3B Instruct
llm_load_print_meta: BOS token = 128000 '<|begin_of_text|>'
llm_load_print_meta: EOS token = 128009 '<|eot_id|>'
llm_load_print_meta: LF token = 128 'Ä'
llm_load_print_meta: EOT token = 128009 '<|eot_id|>'
llm_load_print_meta: max token length = 256
llm_load_tensors: ggml ctx size = 0.12 MiB
llm_load_tensors: CPU buffer size = 1518.09 MiB
.....................................................................
llama_new_context_with_model: n_ctx = 131072
llama_new_context_with_model: n_batch = 2048
llama_new_context_with_model: n_ubatch = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base = 500000.0
llama_new_context_with_model: freq_scale = 1
ggml_backend_cpu_buffer_type_alloc_buffer: failed to allocate buffer of size 15032385568
llama_kv_cache_init: failed to allocate buffer for kv cache
llama_new_context_with_model: llama_kv_cache_init() failed for self-attention cache
llama_init_from_gpt_params: error: failed to create context with model '/home/kevin/data/projects/kg_proj/rvv_transformer/Llama-3.2-3B-Instruct-IQ3_M.gguf'
main: error: unable to load model
可能原因
llama.cpp Github issue： Bug: ggml_backend_cpu_buffer_type_alloc_buffer: failed to allocate buffer of size 137438953504
解决办法
尝试调整模型的参数：
llama.cpp Github参数说明
调整ctx参数后成功运行
kevin@BRICKHOUSE01:~/data/projects/kg_proj/rvv_transformer/llama.cpp$ qemu-riscv64 -L /home/kevin/data/projects/tools/riscv64_linux_gcc/sysroot -cpu rv64,v=true,vlen=256,elen=64,vext_spec=v1.0 ./llama-cli -m /home/kevin/data/projects/kg_proj/rvv_transformer/Llama-3.2-3B-Instruct-IQ3_M.gguf -p "Anything" -n 9 -c 50
Log start
main: build = 3733 (e5701063)
main: built with riscv64-unknown-linux-gnu-gcc () 13.2.0 for riscv64-unknown-linux-gnu
llama_model_loader: loaded meta data with 35 key-value pairs and 255 tensors from /home/kevin/data/projects/kg_proj/rvv_transformer/Llama-3.2-3B-Instruct-IQ3_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv 0: general.architecture str = llama
llama_model_loader: - kv 1: general.type str = model
llama_model_loader: - kv 2: general.name str = Llama 3.2 3B Instruct
llama_model_loader: - kv 3: general.finetune str = Instruct
llama_model_loader: - kv 4: general.basename str = Llama-3.2
llama_model_loader: - kv 5: general.size_label str = 3B
llama_model_loader: - kv 6: general.license str = llama3.2
llama_model_loader: - kv 7: general.tags arr[str,6] = ["facebook", "meta", "pytorch", "llam...
llama_model_loader: - kv 8: general.languages arr[str,8] = ["en", "de", "fr", "it", "pt", "hi", ...
llama_model_loader: - kv 9: llama.block_count u32 = 28
llama_model_loader: - kv 10: llama.context_length u32 = 131072
llama_model_loader: - kv 11: llama.embedding_length u32 = 3072
llama_model_loader: - kv 12: llama.feed_forward_length u32 = 8192
llama_model_loader: - kv 13: llama.attention.head_count u32 = 24
llama_model_loader: - kv 14: llama.attention.head_count_kv u32 = 8
llama_model_loader: - kv 15: llama.rope.freq_base f32 = 500000.000000
llama_model_loader: - kv 16: llama.attention.layer_norm_rms_epsilon f32 = 0.000010
llama_model_loader: - kv 17: llama.attention.key_length u32 = 128
llama_model_loader: - kv 18: llama.attention.value_length u32 = 128
llama_model_loader: - kv 19: general.file_type u32 = 27
llama_model_loader: - kv 20: llama.vocab_size u32 = 128256
llama_model_loader: - kv 21: llama.rope.dimension_count u32 = 128
llama_model_loader: - kv 22: tokenizer.ggml.model str = gpt2
llama_model_loader: - kv 23: tokenizer.ggml.pre str = llama-bpe
llama_model_loader: - kv 24: tokenizer.ggml.tokens arr[str,128256] = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv 25: tokenizer.ggml.token_type arr[i32,128256] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 26: tokenizer.ggml.merges arr[str,280147] = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
llama_model_loader: - kv 27: tokenizer.ggml.bos_token_id u32 = 128000
llama_model_loader: - kv 28: tokenizer.ggml.eos_token_id u32 = 128009
llama_model_loader: - kv 29: tokenizer.chat_template str = {{- bos_token }}\n{%- if custom_tools ...
llama_model_loader: - kv 30: general.quantization_version u32 = 2
llama_model_loader: - kv 31: quantize.imatrix.file str = /models_out/Llama-3.2-3B-Instruct-GGU...
llama_model_loader: - kv 32: quantize.imatrix.dataset str = /training_dir/calibration_datav3.txt
llama_model_loader: - kv 33: quantize.imatrix.entries_count i32 = 196
llama_model_loader: - kv 34: quantize.imatrix.chunks_count i32 = 125
llama_model_loader: - type f32: 58 tensors
llama_model_loader: - type q4_K: 59 tensors
llama_model_loader: - type q6_K: 1 tensors
llama_model_loader: - type iq3_s: 137 tensors
llm_load_vocab: special tokens cache size = 256
llm_load_vocab: token to piece cache size = 0.7999 MB
llm_load_print_meta: format = GGUF V3 (latest)
llm_load_print_meta: arch = llama
llm_load_print_meta: vocab type = BPE
llm_load_print_meta: n_vocab = 128256
llm_load_print_meta: n_merges = 280147
llm_load_print_meta: vocab_only = 0
llm_load_print_meta: n_ctx_train = 131072
llm_load_print_meta: n_embd = 3072
llm_load_print_meta: n_layer = 28
llm_load_print_meta: n_head = 24
llm_load_print_meta: n_head_kv = 8
llm_load_print_meta: n_rot = 128
llm_load_print_meta: n_swa = 0
llm_load_print_meta: n_embd_head_k = 128
llm_load_print_meta: n_embd_head_v = 128
llm_load_print_meta: n_gqa = 3
llm_load_print_meta: n_embd_k_gqa = 1024
llm_load_print_meta: n_embd_v_gqa = 1024
llm_load_print_meta: f_norm_eps = 0.0e+00
llm_load_print_meta: f_norm_rms_eps = 1.0e-05
llm_load_print_meta: f_clamp_kqv = 0.0e+00
llm_load_print_meta: f_max_alibi_bias = 0.0e+00
llm_load_print_meta: f_logit_scale = 0.0e+00
llm_load_print_meta: n_ff = 8192
llm_load_print_meta: n_expert = 0
llm_load_print_meta: n_expert_used = 0
llm_load_print_meta: causal attn = 1
llm_load_print_meta: pooling type = 0
llm_load_print_meta: rope type = 0
llm_load_print_meta: rope scaling = linear
llm_load_print_meta: freq_base_train = 500000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn = 131072
llm_load_print_meta: rope_finetuned = unknown
llm_load_print_meta: ssm_d_conv = 0
llm_load_print_meta: ssm_d_inner = 0
llm_load_print_meta: ssm_d_state = 0
llm_load_print_meta: ssm_dt_rank = 0
llm_load_print_meta: ssm_dt_b_c_rms = 0
llm_load_print_meta: model type = ?B
llm_load_print_meta: model ftype = IQ3_S mix - 3.66 bpw
llm_load_print_meta: model params = 3.21 B
llm_load_print_meta: model size = 1.48 GiB (3.96 BPW)
llm_load_print_meta: general.name = Llama 3.2 3B Instruct
llm_load_print_meta: BOS token = 128000 '<|begin_of_text|>'
llm_load_print_meta: EOS token = 128009 '<|eot_id|>'
llm_load_print_meta: LF token = 128 'Ä'
llm_load_print_meta: EOT token = 128009 '<|eot_id|>'
llm_load_print_meta: max token length = 256
llm_load_tensors: ggml ctx size = 0.12 MiB
llm_load_tensors: CPU buffer size = 1518.09 MiB
.....................................................................
llama_new_context_with_model: n_ctx = 64
llama_new_context_with_model: n_batch = 64
llama_new_context_with_model: n_ubatch = 64
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base = 500000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init: CPU KV buffer size = 7.00 MiB
llama_new_context_with_model: KV self size = 7.00 MiB, K (f16): 3.50 MiB, V (f16): 3.50 MiB
llama_new_context_with_model: CPU output buffer size = 0.49 MiB
llama_new_context_with_model: CPU compute buffer size = 32.06 MiB
llama_new_context_with_model: graph nodes = 902
llama_new_context_with_model: graph splits = 1
system_info: n_threads = 6 (n_threads_batch = 6) / 12 | AVX = 0 | AVX_VNNI = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 0 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 0 | FP16_VA = 0 | RISCV_VECT = 1 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 0 | SSSE3 = 0 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 |
sampling seed: 2622092847
sampling params:
repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
top_k = 40, tfs_z = 1.000, top_p = 0.950, min_p = 0.050, typical_p = 1.000, temp = 0.800
mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler constr:
logits -> logit-bias -> penalties -> top-k -> tail-free -> typical -> top-p -> min-p -> temp-ext -> softmax -> dist
generate: n_ctx = 64, n_batch = 2048, n_predict = 9, n_keep = 1
Anything else?
Yes. You can also consider the
llama_perf_print: sampling time = 10.96 ms / 11 runs ( 1.00 ms per token, 1003.37 tokens per second)
llama_perf_print: load time = 73770.30 ms
llama_perf_print: prompt eval time = 15435.63 ms / 2 tokens ( 7717.81 ms per token, 0.13 tokens per second)
llama_perf_print: eval time = 74267.24 ms / 8 runs ( 9283.41 ms per token, 0.11 tokens per second)
llama_perf_print: total time = 89764.85 ms / 10 tokens
Log end
2024/10/05：生成非向量支持的riscv版本llama.cpp，进行对比实验
llama.cpp编译
make CC="riscv64-unknown-linux-gnu-gcc -march=rv64gc -mabi=lp64d" CXX="riscv64-unknown-linux-gnu-g++ -march=rv64gc -mabi=lp64d"
make CC="riscv64-unknown-linux-gnu-gcc -march=rv64gc -mabi=lp64d" CXX="riscv64-unknown-linux-gnu-g++ -march=rv64gc -mabi=lp64d"
报错：riscv64-unknown-linux-gnu-g++: error: ‘-march=native’: ISA string must begin with rv32 or rv64
解决方案：
GCC GitHub：riscv64-unknown-linux-gnu-g++: error: ‘-march=native’: ISA string must begin with rv32 or rv64
即将llama.cpp工程中的Makefile中523和524行修改为rv64gc编译，该部分原先是交叉编译时使能RVV编译，如图所示：
修正后编译命令
make RISCV_CROSS_COMPILE=1
运行结果
kevin@BRICKHOUSE01:~/data/projects/kg_proj/rvv_transformer/llama.cpp.rv64gc$ qemu-riscv64 -L /home/kevin/data/projects/tools/riscv64_linux_gcc/sysroot -cpu rv64 ./llama-cli -m /home/kevin/data/projects/kg_proj/rvv_transformer/codellama-7b.Q4_K_M.gguf -p "Anything" -n 100 -c 1024
Log start
main: build = 3733 (e5701063)
main: built with riscv64-unknown-linux-gnu-gcc () 13.2.0 for riscv64-unknown-linux-gnu
llama_model_loader: loaded meta data with 20 key-value pairs and 291 tensors from /home/kevin/data/projects/kg_proj/rvv_transformer/codellama-7b.Q4_K_M.gguf (version GGUF V2)
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv 0: general.architecture str = llama
llama_model_loader: - kv 1: general.name str = codellama_codellama-7b-hf
llama_model_loader: - kv 2: llama.context_length u32 = 16384
llama_model_loader: - kv 3: llama.embedding_length u32 = 4096
llama_model_loader: - kv 4: llama.block_count u32 = 32
llama_model_loader: - kv 5: llama.feed_forward_length u32 = 11008
llama_model_loader: - kv 6: llama.rope.dimension_count u32 = 128
llama_model_loader: - kv 7: llama.attention.head_count u32 = 32
llama_model_loader: - kv 8: llama.attention.head_count_kv u32 = 32
llama_model_loader: - kv 9: llama.attention.layer_norm_rms_epsilon f32 = 0.000010
llama_model_loader: - kv 10: llama.rope.freq_base f32 = 1000000.000000
llama_model_loader: - kv 11: general.file_type u32 = 15
llama_model_loader: - kv 12: tokenizer.ggml.model str = llama
llama_model_loader: - kv 13: tokenizer.ggml.tokens arr[str,32016] = ["<unk>", "<s>", "</s>", "<0x00>", "<...
llama_model_loader: - kv 14: tokenizer.ggml.scores arr[f32,32016] = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv 15: tokenizer.ggml.token_type arr[i32,32016] = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...
llama_model_loader: - kv 16: tokenizer.ggml.bos_token_id u32 = 1
llama_model_loader: - kv 17: tokenizer.ggml.eos_token_id u32 = 2
llama_model_loader: - kv 18: tokenizer.ggml.unknown_token_id u32 = 0
llama_model_loader: - kv 19: general.quantization_version u32 = 2
llama_model_loader: - type f32: 65 tensors
llama_model_loader: - type q4_K: 193 tensors
llama_model_loader: - type q6_K: 33 tensors
llm_load_vocab: special tokens cache size = 3
llm_load_vocab: token to piece cache size = 0.1686 MB
llm_load_print_meta: format = GGUF V2
llm_load_print_meta: arch = llama
llm_load_print_meta: vocab type = SPM
llm_load_print_meta: n_vocab = 32016
llm_load_print_meta: n_merges = 0
llm_load_print_meta: vocab_only = 0
llm_load_print_meta: n_ctx_train = 16384
llm_load_print_meta: n_embd = 4096
llm_load_print_meta: n_layer = 32
llm_load_print_meta: n_head = 32
llm_load_print_meta: n_head_kv = 32
llm_load_print_meta: n_rot = 128
llm_load_print_meta: n_swa = 0
llm_load_print_meta: n_embd_head_k = 128
llm_load_print_meta: n_embd_head_v = 128
llm_load_print_meta: n_gqa = 1
llm_load_print_meta: n_embd_k_gqa = 4096
llm_load_print_meta: n_embd_v_gqa = 4096
llm_load_print_meta: f_norm_eps = 0.0e+00
llm_load_print_meta: f_norm_rms_eps = 1.0e-05
llm_load_print_meta: f_clamp_kqv = 0.0e+00
llm_load_print_meta: f_max_alibi_bias = 0.0e+00
llm_load_print_meta: f_logit_scale = 0.0e+00
llm_load_print_meta: n_ff = 11008
llm_load_print_meta: n_expert = 0
llm_load_print_meta: n_expert_used = 0
llm_load_print_meta: causal attn = 1
llm_load_print_meta: pooling type = 0
llm_load_print_meta: rope type = 0
llm_load_print_meta: rope scaling = linear
llm_load_print_meta: freq_base_train = 1000000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn = 16384
llm_load_print_meta: rope_finetuned = unknown
llm_load_print_meta: ssm_d_conv = 0
llm_load_print_meta: ssm_d_inner = 0
llm_load_print_meta: ssm_d_state = 0
llm_load_print_meta: ssm_dt_rank = 0
llm_load_print_meta: ssm_dt_b_c_rms = 0
llm_load_print_meta: model type = 7B
llm_load_print_meta: model ftype = Q4_K - Medium
llm_load_print_meta: model params = 6.74 B
llm_load_print_meta: model size = 3.80 GiB (4.84 BPW)
llm_load_print_meta: general.name = codellama_codellama-7b-hf
llm_load_print_meta: BOS token = 1 '<s>'
llm_load_print_meta: EOS token = 2 '</s>'
llm_load_print_meta: UNK token = 0 '<unk>'
llm_load_print_meta: LF token = 13 '<0x0A>'
llm_load_print_meta: PRE token = 32007 '▁<PRE>'
llm_load_print_meta: SUF token = 32008 '▁<SUF>'
llm_load_print_meta: MID token = 32009 '▁<MID>'
llm_load_print_meta: EOT token = 32010 '▁<EOT>'
llm_load_print_meta: max token length = 48
llm_load_tensors: ggml ctx size = 0.14 MiB
llm_load_tensors: CPU buffer size = 3891.33 MiB
..................................................................................................
llama_new_context_with_model: n_ctx = 1024
llama_new_context_with_model: n_batch = 1024
llama_new_context_with_model: n_ubatch = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base = 1000000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init: CPU KV buffer size = 512.00 MiB
llama_new_context_with_model: KV self size = 512.00 MiB, K (f16): 256.00 MiB, V (f16): 256.00 MiB
llama_new_context_with_model: CPU output buffer size = 0.12 MiB
llama_new_context_with_model: CPU compute buffer size = 98.01 MiB
llama_new_context_with_model: graph nodes = 1030
llama_new_context_with_model: graph splits = 1
system_info: n_threads = 6 (n_threads_batch = 6) / 12 | AVX = 0 | AVX_VNNI = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 0 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 0 | FP16_VA = 0 | RISCV_VECT = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 0 | SSSE3 = 0 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 |
sampling seed: 3630221793
sampling params:
repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
top_k = 40, tfs_z = 1.000, top_p = 0.950, min_p = 0.050, typical_p = 1.000, temp = 0.800
mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler constr:
logits -> logit-bias -> penalties -> top-k -> tail-free -> typical -> top-p -> min-p -> temp-ext -> softmax -> dist
generate: n_ctx = 1024, n_batch = 2048, n_predict = 100, n_keep = 1
Anything is possible if you believe it. The only thing you are not capable of doing is believing in anything.
- John L. Sullivan
- To be honest, I've never had a problem with the fact that I'm a bad person and I'm a horrible human. I've always been okay with that.
- I think it's a mistake to think of yourself as being in the middle of the world, instead of at the
llama_perf_print: sampling time = 94.53 ms / 104 runs ( 0.91 ms per token, 1100.12 tokens per second)
llama_perf_print: load time = 44296.57 ms
llama_perf_print: prompt eval time = 79786.77 ms / 4 tokens (19946.69 ms per token, 0.05 tokens per second)
llama_perf_print: eval time = 2360379.06 ms / 99 runs (23842.21 ms per token, 0.04 tokens per second)
llama_perf_print: total time = 2440911.31 ms / 103 tokens
Log end
