{ "cells": [ { "cell_type": "markdown", "id": "0d641747", "metadata": {}, "source": [ "## Llama CPP\n", "\n", "Lecture 19 | CMU ANLP Fall 2025 | Instructor: Sean Welleck\n", "\n", "This notebook shows running inference with a quantized Qwen3-8B via the llama cpp server.\n", "\n", "After running the cells below, navigate to http://localhost:8080/ to interact with the model." ] }, { "cell_type": "code", "execution_count": null, "id": "70d8ee53", "metadata": {}, "outputs": [], "source": [ "!brew install llama.cpp" ] }, { "cell_type": "code", "execution_count": 8, "id": "df3de2db", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "ggml_metal_device_init: tensor API disabled for pre-M5 device\n", "ggml_metal_library_init: using embedded metal library\n", "ggml_metal_library_init: loaded in 0.022 sec\n", "ggml_metal_device_init: GPU name: Apple M4\n", "ggml_metal_device_init: GPU family: MTLGPUFamilyApple9 (1009)\n", "ggml_metal_device_init: GPU family: MTLGPUFamilyCommon3 (3003)\n", "ggml_metal_device_init: GPU family: MTLGPUFamilyMetal3 (5001)\n", "ggml_metal_device_init: simdgroup reduction = true\n", "ggml_metal_device_init: simdgroup matrix mul. = true\n", "ggml_metal_device_init: has unified memory = true\n", "ggml_metal_device_init: has bfloat = true\n", "ggml_metal_device_init: has tensor = false\n", "ggml_metal_device_init: use residency sets = true\n", "ggml_metal_device_init: use shared buffers = true\n", "ggml_metal_device_init: recommendedMaxWorkingSetSize = 11453.25 MB\n", "* Host huggingface.co:443 was resolved.\n", "* IPv6: (none)\n", "* IPv4: 3.167.112.38, 3.167.112.96, 3.167.112.45, 3.167.112.25\n", "* Trying 3.167.112.38:443...\n", "* Connected to huggingface.co (3.167.112.38) port 443\n", "* ALPN: curl offers h2,http/1.1\n", "* CAfile: /etc/ssl/cert.pem\n", "* CApath: none\n", "* SSL connection using TLSv1.3 / AEAD-AES128-GCM-SHA256 / [blank] / UNDEF\n", "* ALPN: server accepted h2\n", "* Server certificate:\n", "* subject: CN=huggingface.co\n", "* start date: Apr 13 00:00:00 2025 GMT\n", "* expire date: May 12 23:59:59 2026 GMT\n", "* subjectAltName: host \"huggingface.co\" matched cert's \"huggingface.co\"\n", "* issuer: C=US; O=Amazon; CN=Amazon RSA 2048 M02\n", "* SSL certificate verify ok.\n", "* using HTTP/2\n", "* [HTTP/2] [1] OPENED stream for https://huggingface.co/v2/Qwen/Qwen3-8B-GGUF/manifests/Q4_K_M\n", "* [HTTP/2] [1] [:method: GET]\n", "* [HTTP/2] [1] [:scheme: https]\n", "* [HTTP/2] [1] [:authority: huggingface.co]\n", "* [HTTP/2] [1] [:path: /v2/Qwen/Qwen3-8B-GGUF/manifests/Q4_K_M]\n", "* [HTTP/2] [1] [user-agent: llama-cpp]\n", "* [HTTP/2] [1] [accept: application/json]\n", "> GET /v2/Qwen/Qwen3-8B-GGUF/manifests/Q4_K_M HTTP/2\n", "Host: huggingface.co\n", "User-Agent: llama-cpp\n", "Accept: application/json\n", "\n", "* Request completely sent off\n", "< HTTP/2 200 \n", "< content-type: application/json; charset=utf-8\n", "< content-length: 1128\n", "< date: Tue, 11 Nov 2025 19:59:34 GMT\n", "< etag: W/\"468-s23Zs47Cv1pa92kkfI/4vBNEXPc\"\n", "< x-powered-by: huggingface-moon\n", "< x-request-id: Root=1-691395a6-0f5453d51799ed551ef50b10\n", "< ratelimit: \"pages\";r=99;t=194\n", "< ratelimit-policy: \"fixed window\";\"pages\";q=100;w=300\n", "< cross-origin-opener-policy: same-origin\n", "< referrer-policy: strict-origin-when-cross-origin\n", "< access-control-max-age: 86400\n", "< access-control-allow-origin: https://huggingface.co\n", "< vary: Origin\n", "< access-control-expose-headers: X-Repo-Commit,X-Request-Id,X-Error-Code,X-Error-Message,X-Total-Count,ETag,Link,Accept-Ranges,Content-Range,X-Linked-Size,X-Linked-ETag,X-Xet-Hash\n", "< x-cache: Miss from cloudfront\n", "< via: 1.1 7ba8238566c8641cae9d38ebcb67e0e8.cloudfront.net (CloudFront)\n", "< x-amz-cf-pop: IAD55-P8\n", "< x-amz-cf-id: O4purX4ZXYrcIDZWgw8sJJEXNWWCjOEQMWH11YHYjS15HVxd6-lu8A==\n", "< \n", "* Connection #0 to host huggingface.co left intact\n", "common_download_file_single_online: using cached file: /Users/wellecks/Library/Caches/llama.cpp/Qwen_Qwen3-8B-GGUF_Qwen3-8B-Q4_K_M.gguf\n", "main: setting n_parallel = 4 and kv_unified = true (add -kvu to disable this)\n", "\u001b[0mbuild: 6990 (53d7d21e6) with Apple clang version 17.0.0 (clang-1700.4.4.1) for arm64-apple-darwin24.6.0\n", "system info: n_threads = 4, n_threads_batch = 4, total_threads = 10\n", "\n", "system_info: n_threads = 4 (n_threads_batch = 4) / 10 | Metal : EMBED_LIBRARY = 1 | CPU : NEON = 1 | ARM_FMA = 1 | FP16_VA = 1 | DOTPROD = 1 | LLAMAFILE = 1 | ACCELERATE = 1 | REPACK = 1 | \n", "\n", "main: binding port with default address family\n", "main: HTTP server is listening, hostname: 127.0.0.1, port: 8080, http threads: 9\n", "main: loading model\n", "srv load_model: loading model '/Users/wellecks/Library/Caches/llama.cpp/Qwen_Qwen3-8B-GGUF_Qwen3-8B-Q4_K_M.gguf'\n", "llama_model_load_from_file_impl: using device Metal (Apple M4) (unknown id) - 10922 MiB free\n", "llama_model_loader: loaded meta data with 28 key-value pairs and 399 tensors from /Users/wellecks/Library/Caches/llama.cpp/Qwen_Qwen3-8B-GGUF_Qwen3-8B-Q4_K_M.gguf (version GGUF V3 (latest))\n", "llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", "llama_model_loader: - kv 0: general.architecture str = qwen3\n", "llama_model_loader: - kv 1: general.type str = model\n", "llama_model_loader: - kv 2: general.name str = Qwen3 8B Awq Compatible Instruct\n", "llama_model_loader: - kv 3: general.finetune str = awq-compatible-Instruct\n", "llama_model_loader: - kv 4: general.basename str = Qwen3\n", "llama_model_loader: - kv 5: general.size_label str = 8B\n", "llama_model_loader: - kv 6: qwen3.block_count u32 = 36\n", "llama_model_loader: - kv 7: qwen3.context_length u32 = 40960\n", "llama_model_loader: - kv 8: qwen3.embedding_length u32 = 4096\n", "llama_model_loader: - kv 9: qwen3.feed_forward_length u32 = 12288\n", "llama_model_loader: - kv 10: qwen3.attention.head_count u32 = 32\n", "llama_model_loader: - kv 11: qwen3.attention.head_count_kv u32 = 8\n", "llama_model_loader: - kv 12: qwen3.rope.freq_base f32 = 1000000.000000\n", "llama_model_loader: - kv 13: qwen3.attention.layer_norm_rms_epsilon f32 = 0.000001\n", "llama_model_loader: - kv 14: qwen3.attention.key_length u32 = 128\n", "llama_model_loader: - kv 15: qwen3.attention.value_length u32 = 128\n", "llama_model_loader: - kv 16: tokenizer.ggml.model str = gpt2\n", "llama_model_loader: - kv 17: tokenizer.ggml.pre str = qwen2\n", "llama_model_loader: - kv 18: tokenizer.ggml.tokens arr[str,151936] = [\"!\", \"\\\"\", \"#\", \"$\", \"%\", \"&\", \"'\", ...\n", "llama_model_loader: - kv 19: tokenizer.ggml.token_type arr[i32,151936] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...\n", "llama_model_loader: - kv 20: tokenizer.ggml.merges arr[str,151387] = [\"Ġ Ġ\", \"ĠĠ ĠĠ\", \"i n\", \"Ġ t\",...\n", "llama_model_loader: - kv 21: tokenizer.ggml.eos_token_id u32 = 151645\n", "llama_model_loader: - kv 22: tokenizer.ggml.padding_token_id u32 = 151643\n", "llama_model_loader: - kv 23: tokenizer.ggml.bos_token_id u32 = 151643\n", "llama_model_loader: - kv 24: tokenizer.ggml.add_bos_token bool = false\n", "llama_model_loader: - kv 25: tokenizer.chat_template str = {%- if tools %}\\n {{- '<|im_start|>...\n", "llama_model_loader: - kv 26: general.quantization_version u32 = 2\n", "llama_model_loader: - kv 27: general.file_type u32 = 15\n", "llama_model_loader: - type f32: 145 tensors\n", "llama_model_loader: - type q4_K: 217 tensors\n", "llama_model_loader: - type q6_K: 37 tensors\n", "print_info: file format = GGUF V3 (latest)\n", "print_info: file type = Q4_K - Medium\n", "print_info: file size = 4.68 GiB (4.90 BPW) \n", "load: printing all EOG tokens:\n", "load: - 151643 ('<|endoftext|>')\n", "load: - 151645 ('<|im_end|>')\n", "load: - 151662 ('<|fim_pad|>')\n", "load: - 151663 ('<|repo_name|>')\n", "load: - 151664 ('<|file_sep|>')\n", "load: special tokens cache size = 26\n", "load: token to piece cache size = 0.9311 MB\n", "print_info: arch = qwen3\n", "print_info: vocab_only = 0\n", "print_info: n_ctx_train = 40960\n", "print_info: n_embd = 4096\n", "print_info: n_embd_inp = 4096\n", "print_info: n_layer = 36\n", "print_info: n_head = 32\n", "print_info: n_head_kv = 8\n", "print_info: n_rot = 128\n", "print_info: n_swa = 0\n", "print_info: is_swa_any = 0\n", "print_info: n_embd_head_k = 128\n", "print_info: n_embd_head_v = 128\n", "print_info: n_gqa = 4\n", "print_info: n_embd_k_gqa = 1024\n", "print_info: n_embd_v_gqa = 1024\n", "print_info: f_norm_eps = 0.0e+00\n", "print_info: f_norm_rms_eps = 1.0e-06\n", "print_info: f_clamp_kqv = 0.0e+00\n", "print_info: f_max_alibi_bias = 0.0e+00\n", "print_info: f_logit_scale = 0.0e+00\n", "print_info: f_attn_scale = 0.0e+00\n", "print_info: n_ff = 12288\n", "print_info: n_expert = 0\n", "print_info: n_expert_used = 0\n", "print_info: n_expert_groups = 0\n", "print_info: n_group_used = 0\n", "print_info: causal attn = 1\n", "print_info: pooling type = -1\n", "print_info: rope type = 2\n", "print_info: rope scaling = linear\n", "print_info: freq_base_train = 1000000.0\n", "print_info: freq_scale_train = 1\n", "print_info: n_ctx_orig_yarn = 40960\n", "print_info: rope_finetuned = unknown\n", "print_info: model type = 8B\n", "print_info: model params = 8.19 B\n", "print_info: general.name = Qwen3 8B Awq Compatible Instruct\n", "print_info: vocab type = BPE\n", "print_info: n_vocab = 151936\n", "print_info: n_merges = 151387\n", "print_info: BOS token = 151643 '<|endoftext|>'\n", "print_info: EOS token = 151645 '<|im_end|>'\n", "print_info: EOT token = 151645 '<|im_end|>'\n", "print_info: PAD token = 151643 '<|endoftext|>'\n", "print_info: LF token = 198 'Ċ'\n", "print_info: FIM PRE token = 151659 '<|fim_prefix|>'\n", "print_info: FIM SUF token = 151661 '<|fim_suffix|>'\n", "print_info: FIM MID token = 151660 '<|fim_middle|>'\n", "print_info: FIM PAD token = 151662 '<|fim_pad|>'\n", "print_info: FIM REP token = 151663 '<|repo_name|>'\n", "print_info: FIM SEP token = 151664 '<|file_sep|>'\n", "print_info: EOG token = 151643 '<|endoftext|>'\n", "print_info: EOG token = 151645 '<|im_end|>'\n", "print_info: EOG token = 151662 '<|fim_pad|>'\n", "print_info: EOG token = 151663 '<|repo_name|>'\n", "print_info: EOG token = 151664 '<|file_sep|>'\n", "print_info: max token length = 256\n", "load_tensors: loading model tensors, this can take a while... (mmap = true)\n", "load_tensors: offloading 36 repeating layers to GPU\n", "load_tensors: offloading output layer to GPU\n", "load_tensors: offloaded 37/37 layers to GPU\n", "load_tensors: CPU_Mapped model buffer size = 333.84 MiB\n", "load_tensors: Metal_Mapped model buffer size = 4789.19 MiB\n", ".....................................................................................\n", "llama_context: constructing llama_context\n", "llama_context: n_seq_max = 4\n", "llama_context: n_ctx = 4096\n", "llama_context: n_ctx_seq = 4096\n", "llama_context: n_batch = 2048\n", "llama_context: n_ubatch = 512\n", "llama_context: causal_attn = 1\n", "llama_context: flash_attn = auto\n", "llama_context: kv_unified = true\n", "llama_context: freq_base = 1000000.0\n", "llama_context: freq_scale = 1\n", "llama_context: n_ctx_seq (4096) < n_ctx_train (40960) -- the full capacity of the model will not be utilized\n", "\u001b[0mggml_metal_init: allocating\n", "ggml_metal_init: found device: Apple M4\n", "ggml_metal_init: picking default device: Apple M4\n", "ggml_metal_init: use fusion = true\n", "ggml_metal_init: use concurrency = true\n", "ggml_metal_init: use graph optimize = true\n", "llama_context: CPU output buffer size = 2.32 MiB\n", "llama_kv_cache: Metal KV buffer size = 576.00 MiB\n", "llama_kv_cache: size = 576.00 MiB ( 4096 cells, 36 layers, 4/1 seqs), K (f16): 288.00 MiB, V (f16): 288.00 MiB\n", "llama_context: Flash Attention was auto, set to enabled\n", "llama_context: Metal compute buffer size = 304.75 MiB\n", "llama_context: CPU compute buffer size = 16.01 MiB\n", "llama_context: graph nodes = 1267\n", "llama_context: graph splits = 2\n", "common_init_from_params: added <|endoftext|> logit bias = -inf\n", "common_init_from_params: added <|im_end|> logit bias = -inf\n", "common_init_from_params: added <|fim_pad|> logit bias = -inf\n", "common_init_from_params: added <|repo_name|> logit bias = -inf\n", "common_init_from_params: added <|file_sep|> logit bias = -inf\n", "common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096\n", "common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", "\u001b[0msrv init: initializing slots, n_slots = 4\n", "slot init: id 0 | task -1 | new slot, n_ctx = 4096\n", "slot init: id 1 | task -1 | new slot, n_ctx = 4096\n", "slot init: id 2 | task -1 | new slot, n_ctx = 4096\n", "slot init: id 3 | task -1 | new slot, n_ctx = 4096\n", "srv init: prompt cache is enabled, size limit: 8192 MiB\n", "\u001b[0msrv init: use `--cache-ram 0` to disable the prompt cache\n", "\u001b[0msrv init: for more info see https://github.com/ggml-org/llama.cpp/pull/16391\n", "\u001b[0msrv init: thinking = 0\n", "main: model loaded\n", "main: chat template, chat_template: {%- if tools %}\n", " {{- '<|im_start|>system\\n' }}\n", " {%- if messages[0].role == 'system' %}\n", " {{- messages[0].content + '\\n\\n' }}\n", " {%- endif %}\n", " {{- \"# Tools\\n\\nYou may call one or more functions to assist with the user query.\\n\\nYou are provided with function signatures within XML tags:\\n\" }}\n", " {%- for tool in tools %}\n", " {{- \"\\n\" }}\n", " {{- tool | tojson }}\n", " {%- endfor %}\n", " {{- \"\\n\\n\\nFor each function call, return a json object with function name and arguments within XML tags:\\n\\n{\\\"name\\\": , \\\"arguments\\\": }\\n<|im_end|>\\n\" }}\n", "{%- else %}\n", " {%- if messages[0].role == 'system' %}\n", " {{- '<|im_start|>system\\n' + messages[0].content + '<|im_end|>\\n' }}\n", " {%- endif %}\n", "{%- endif %}\n", "{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}\n", "{%- for index in range(ns.last_query_index, -1, -1) %}\n", " {%- set message = messages[index] %}\n", " {%- if ns.multi_step_tool and message.role == \"user\" and not('' in message.content and '' in message.content) %}\n", " {%- set ns.multi_step_tool = false %}\n", " {%- set ns.last_query_index = index %}\n", " {%- endif %}\n", "{%- endfor %}\n", "{%- for message in messages %}\n", " {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) %}\n", " {{- '<|im_start|>' + message.role + '\\n' + message.content + '<|im_end|>' + '\\n' }}\n", " {%- elif message.role == \"assistant\" %}\n", " {%- set content = message.content %}\n", " {%- set reasoning_content = '' %}\n", " {%- if message.reasoning_content is defined and message.reasoning_content is not none %}\n", " {%- set reasoning_content = message.reasoning_content %}\n", " {%- else %}\n", " {%- if '' in message.content %}\n", " {%- set content = message.content.split('')[-1].lstrip('\\n') %}\n", " {%- set reasoning_content = message.content.split('')[0].rstrip('\\n').split('')[-1].lstrip('\\n') %}\n", " {%- endif %}\n", " {%- endif %}\n", " {%- if loop.index0 > ns.last_query_index %}\n", " {%- if loop.last or (not loop.last and reasoning_content) %}\n", " {{- '<|im_start|>' + message.role + '\\n\\n' + reasoning_content.strip('\\n') + '\\n\\n\\n' + content.lstrip('\\n') }}\n", " {%- else %}\n", " {{- '<|im_start|>' + message.role + '\\n' + content }}\n", " {%- endif %}\n", " {%- else %}\n", " {{- '<|im_start|>' + message.role + '\\n' + content }}\n", " {%- endif %}\n", " {%- if message.tool_calls %}\n", " {%- for tool_call in message.tool_calls %}\n", " {%- if (loop.first and content) or (not loop.first) %}\n", " {{- '\\n' }}\n", " {%- endif %}\n", " {%- if tool_call.function %}\n", " {%- set tool_call = tool_call.function %}\n", " {%- endif %}\n", " {{- '\\n{\"name\": \"' }}\n", " {{- tool_call.name }}\n", " {{- '\", \"arguments\": ' }}\n", " {%- if tool_call.arguments is string %}\n", " {{- tool_call.arguments }}\n", " {%- else %}\n", " {{- tool_call.arguments | tojson }}\n", " {%- endif %}\n", " {{- '}\\n' }}\n", " {%- endfor %}\n", " {%- endif %}\n", " {{- '<|im_end|>\\n' }}\n", " {%- elif message.role == \"tool\" %}\n", " {%- if loop.first or (messages[loop.index0 - 1].role != \"tool\") %}\n", " {{- '<|im_start|>user' }}\n", " {%- endif %}\n", " {{- '\\n\\n' }}\n", " {{- message.content }}\n", " {{- '\\n' }}\n", " {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n", " {{- '<|im_end|>\\n' }}\n", " {%- endif %}\n", " {%- endif %}\n", "{%- endfor %}\n", "{%- if add_generation_prompt %}\n", " {{- '<|im_start|>assistant\\n' }}\n", " {%- if enable_thinking is defined and enable_thinking is false %}\n", " {{- '\\n\\n\\n\\n' }}\n", " {%- endif %}\n", "{%- endif %}, example_format: '<|im_start|>system\n", "You are a helpful assistant<|im_end|>\n", "<|im_start|>user\n", "Hello<|im_end|>\n", "<|im_start|>assistant\n", "Hi there<|im_end|>\n", "<|im_start|>user\n", "How are you?<|im_end|>\n", "<|im_start|>assistant\n", "'\n", "main: server is listening on http://127.0.0.1:8080 - starting the main loop\n", "srv update_slots: all slots are idle\n", "srv log_server_r: request: GET / 127.0.0.1 200\n", "srv log_server_r: request: GET /props 127.0.0.1 200\n", "srv log_server_r: request: GET /props 127.0.0.1 200\n", "srv log_server_r: request: GET /props 127.0.0.1 200\n", "srv update_slots: all slots are idle\n", "srv log_server_r: request: GET /slots 127.0.0.1 200\n", "srv log_server_r: request: GET /props 127.0.0.1 200\n", "srv log_server_r: request: GET /props 127.0.0.1 200\n", "srv params_from_: Chat format: Content-only\n", "slot get_availabl: id 3 | task -1 | selected slot by LRU, t_last = -1\n", "slot launch_slot_: id 3 | task -1 | sampler chain: logits -> logit-bias -> penalties -> dry -> top-n-sigma -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist \n", "slot launch_slot_: id 3 | task 1 | processing task\n", "slot update_slots: id 3 | task 1 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 17\n", "slot update_slots: id 3 | task 1 | n_tokens = 0, memory_seq_rm [0, end)\n", "slot update_slots: id 3 | task 1 | prompt processing progress, n_tokens = 17, batch.n_tokens = 17, progress = 1.000000\n", "slot update_slots: id 3 | task 1 | prompt done, n_tokens = 17, batch.n_tokens = 17\n", "srv log_server_r: request: GET /props 127.0.0.1 200\n", "srv log_server_r: request: GET /slots 127.0.0.1 200\n", "srv log_server_r: request: GET /slots 127.0.0.1 200\n", "slot print_timing: id 3 | task 1 | \n", "prompt eval time = 230.02 ms / 17 tokens ( 13.53 ms per token, 73.91 tokens per second)\n", " eval time = 102383.24 ms / 1837 tokens ( 55.73 ms per token, 17.94 tokens per second)\n", " total time = 102613.26 ms / 1854 tokens\n", "slot release: id 3 | task 1 | stop processing: n_tokens = 1853, truncated = 0\n", "srv update_slots: all slots are idle\n", "srv log_server_r: request: POST /v1/chat/completions 127.0.0.1 200\n", "^C\n", "Received second interrupt, terminating immediately.\n" ] } ], "source": [ "!llama-server -hf Qwen/Qwen3-8B-GGUF:Q4_K_M --port 8080" ] }, { "cell_type": "code", "execution_count": null, "id": "682f138b", "metadata": {}, "outputs": [], "source": [] } ], "metadata": { "kernelspec": { "display_name": "anlp", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.10.18" } }, "nbformat": 4, "nbformat_minor": 5 }