K230-Llama3

README.md

@@ -1,3 +1,74 @@
 # rvspoc-S2422-Llama3
+##Llama 3 在 K230 上的优化实现
+Llama3模型模型规模较大，使用SD卡，使用swap的方式，成功在canmv_K230上运行Llama3(Q2，Q4和Q8量化)。
+###############################环境准备######################################
+##烧写镜像
+使用嘉楠开发者社区提供镜像,镜像下载链接：https://developer.canaan-creative.com/resource， 镜像版本：canmv_debian_sdcard_sdk_1.3.img.gz
+
+##安装软件
+为了便于文件传输和重新分区，使用apt安装ssh和parted软件。注意在apt update前需要先同步一下系统时间。
+参考命令：
+1）更新时间：
+date --set="2024-08-03 03:15:20" ##修改为UTC当前时间
+2）apt update 
+3）安装ssh：
+apt install openssh-server -y
+4）安装parted：
+apt install parted -y
+
+##重新分区
+
+参考命令：
+1）查看可用空间大小：
+parted   -l /dev/mmcblk1
+2）根据上条命令运行结果的SD卡大小提示设置分区；
+parted /dev/mmcblk1 resizepart 3 62.5G
+resize2fs /dev/mmcblk1p3
+##设置swap
+swap设置为2G
+1）dd if=/dev/zero of=/mnt/swap bs=256M count=8
+2）mkswap /mnt/swap
+3）swapon /mnt/swap
+###################程序运行#########################
+
+试了两个llama3的实现，llama3.c 做了Q8量化，llama.cpp尝试Q2和Q4量化，具体如下：
+
+1）基于https://github.com/jameswdelancey/llama3.c 实现，使用方法一致。需要生成q8量化文件。
+
+##提交内容
+运行需要三个文件：
+llama3_8b_instruct_q80.bin 
+下载地址：https://huggingface.co/Sophia957/llama3_8b_instruct_q80/resolve/main/llama3_8b_instruct_q80.bin
+runq3-k230 （PR上传）
+tokenizer.bin（PR上传）
+
+其他：
+a. 移植过程中，参考上届的llama2题目，在K230上对比了Llama2增加RVV前后的效果，提速明显，但尝试在llama3.c上增加rvv，没有明显提速且运行结果有点问题，于是尝试其他方法提速。
+b. 由于Llama3中有较多的matmul计算，尝试使用Matrix进行提速，下载工具链：Xuantie-900-gcc-linux-6.6.0-glibc-x86_64-V2.10.1，编译Matrix的demo成功，但在K230中运行没有成功，提示指令不支持，后续查询C908资料没有看到包含Matrix相关介绍。
+
+
+2）基于 https://github.com/ggerganov/llama.cpp 尝试了Q2_k和Q4_0
+
+##提交内容
+运行需要文件：
+llama-cli-tune (PR上传)
+model文件：Meta-Llama-3-8B.Q2_K.gguf 或 Meta-Llama-3-8B.Q4_0.gguf
+下载地址：
+https://huggingface.co/QuantFactory/Meta-Llama-3-8B-GGUF/resolve/main/Meta-Llama-3-8B.Q2_K.gguf
+https://huggingface.co/QuantFactory/Meta-Llama-3-8B-GGUF/resolve/main/Meta-Llama-3-8B.Q4_0.gguf
+
+参考运行命令：
+./llama-cli-tune -m ./Meta-Llama-3-8B.Q2_K.gguf -p "Once upon a time, " -n 10   -s 123
+./llama-cli-tune -m ./Meta-Llama-3-8B.Q4_0.gguf -p "Once upon a time, " -n 10   -s 123
+
+其他：
+a. llama.cpp中使用了rvv的intrinsic,但是目前使用 Xuantie-900-gcc-linux-6.6.0-glibc工具链，rvv intrinsic部分没有成功编译进去。
+
+
+重启设备遇到网络启动失败，可运行下面指令：
+networkctl
+ip link set enu1 name eth0
+systemctl restart networking.service
+
 
 Submission repo for S2422. ref: rvspoc.org

cpp-Q2-log.txt

@@ -0,0 +1,124 @@
+root@v:/home/s# ./llama-cli-tune -m ./Meta-Llama-3-8B.Q2_K.gguf -p "Once upon a time" -n 10   -s 123
+Log start
+main: build = 3507 (4b77ea95)
+main: built with riscv64-unknown-linux-gnu-gcc (Xuantie-900 linux-6.6.0 glibc gcc Toolchain V2.10.1 B-20240712) 10.4.0 for riscv64-unknown-linux-gnu
+main: seed  = 123
+llama_model_loader: loaded meta data with 22 key-value pairs and 291 tensors from ./Meta-Llama-3-8B.Q2_K.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.name str              = .
+llama_model_loader: - kv   2:                           llama.vocab_size u32              = 128256
+llama_model_loader: - kv   3:                       llama.context_length u32              = 8192
+llama_model_loader: - kv   4:                     llama.embedding_length u32              = 4096
+llama_model_loader: - kv   5:                          llama.block_count u32              = 32
+llama_model_loader: - kv   6:                  llama.feed_forward_length u32              = 14336
+llama_model_loader: - kv   7:                 llama.rope.dimension_count u32              = 128
+llama_model_loader: - kv   8:                 llama.attention.head_count u32              = 32
+llama_model_loader: - kv   9:              llama.attention.head_count_kv u32              = 8
+llama_model_loader: - kv  10:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
+llama_model_loader: - kv  11:                       llama.rope.freq_base f32              = 500000.000000
+llama_model_loader: - kv  12:                          general.file_type u32              = 10
+llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
+llama_model_loader: - kv  14:                      tokenizer.ggml.tokens arr[str,128256]  = ["!", "\"", "#", "$", "%", "&", "'", ...
+llama_model_loader: - kv  15:                      tokenizer.ggml.scores arr[f32,128256]  = [0.000000, 0.000000, 0.000000, 0.0000...
+llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,128256]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
+llama_model_loader: - kv  17:                      tokenizer.ggml.merges arr[str,280147]  = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
+llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 128000
+llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 128001
+llama_model_loader: - kv  20:                    tokenizer.chat_template str              = {% set loop_messages = messages %}{% ...
+llama_model_loader: - kv  21:               general.quantization_version u32              = 2
+llama_model_loader: - type  f32:   65 tensors
+llama_model_loader: - type q2_K:  129 tensors
+llama_model_loader: - type q3_K:   64 tensors
+llama_model_loader: - type q4_K:   32 tensors
+llama_model_loader: - type q6_K:    1 tensors
+llm_load_vocab: missing pre-tokenizer type, using: 'default'
+llm_load_vocab:
+llm_load_vocab: ************************************
+llm_load_vocab: GENERATION QUALITY WILL BE DEGRADED!
+llm_load_vocab: CONSIDER REGENERATING THE MODEL
+llm_load_vocab: ************************************
+llm_load_vocab:
+llm_load_vocab: special tokens cache size = 256
+llm_load_vocab: token to piece cache size = 0.8000 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      = 8192
+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        = 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            = 4
+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             = 14336
+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  = 8192
+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: model type       = 8B
+llm_load_print_meta: model ftype      = Q2_K - Medium
+llm_load_print_meta: model params     = 8.03 B
+llm_load_print_meta: model size       = 2.95 GiB (3.16 BPW)
+llm_load_print_meta: general.name     = .
+llm_load_print_meta: BOS token        = 128000 '<|begin_of_text|>'
+llm_load_print_meta: EOS token        = 128001 '<|end_of_text|>'
+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 =    0.14 MiB
+llm_load_tensors:        CPU buffer size =  3024.38 MiB
+...................................................................................
+llama_new_context_with_model: n_ctx      = 8192
+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
+llama_kv_cache_init:        CPU KV buffer size =  1024.00 MiB
+llama_new_context_with_model: KV self size  = 1024.00 MiB, K (f16):  512.00 MiB, V (f16):  512.00 MiB
+llama_new_context_with_model:        CPU  output buffer size =     0.49 MiB
+llama_new_context_with_model:        CPU compute buffer size =   560.01 MiB
+llama_new_context_with_model: graph nodes  = 1030
+llama_new_context_with_model: graph splits = 1
+
+system_info: n_threads = 1 / 1 | 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 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 0 | SSSE3 = 0 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 0 |
+sampling:
+        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
+sampling order:
+CFG -> Penalties -> top_k -> tfs_z -> typical_p -> top_p -> min_p -> temperature
+generate: n_ctx = 8192, n_batch = 2048, n_predict = 10, n_keep = 0
+
+
+Once upon a time there was a big rock on the island of K
+llama_print_timings:        load time =  497741.76 ms
+llama_print_timings:      sample time =      11.44 ms /    10 runs   (    1.14 ms per token,   874.36 tokens per second)
+llama_print_timings: prompt eval time =  237352.11 ms /     4 tokens (59338.03 ms per token,     0.02 tokens per second)
+llama_print_timings:        eval time = 1170584.69 ms /     9 runs   (130064.97 ms per token,     0.01 tokens per second)
+llama_print_timings:       total time = 1408539.24 ms /    13 tokens
+Log end