What happened?
I'm running llama.cpp server with docker. My docker compose command set as the following:
services:
llama-cpp:
image: ghcr.io/ggerganov/llama.cpp:server-cuda
command: ["-m", "/models/Qwen2.5-7B-Instruct-Q4_K_M.gguf", "--port", "8000", "--n-gpu-layers", "49", "--no-mmap", "--ctx-size", "131072", "--rope-scaling", "yarn", "--rope-scale", "4", "--flash-attn"]
environment:
LLAMA_ARG_ROPE_SCALING_TYPE: yarn
LLAMA_ARG_ROPE_SCALE: 4Expectation: the rope scaling type should set to yarn and the rope scale should set to 4
Current result: the rope scaling type is still linear and the scale is still 2, which are all the default options
P.S. I even set the environment variables, and the log showed:
2024-11-17 00:33:21 warn: LLAMA_ARG_ROPE_SCALING_TYPE environment variable is set, but will be overwritten by command line argument --rope-scaling
2024-11-17 00:33:21 warn: LLAMA_ARG_ROPE_SCALE environment variable is set, but will be overwritten by command line argument --rope-scale
But somehow things still didn't work
Name and Version
build: 4112 (eda7e1d) with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
What operating system are you seeing the problem on?
Linux, Windows
Relevant log output
2024-11-17 00:33:21 ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
2024-11-17 00:33:21 ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
2024-11-17 00:33:21 ggml_cuda_init: found 1 CUDA devices:
2024-11-17 00:33:21 Device 0: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes
2024-11-17 00:33:21 warn: LLAMA_ARG_ROPE_SCALING_TYPE environment variable is set, but will be overwritten by command line argument --rope-scaling
2024-11-17 00:33:21 warn: LLAMA_ARG_ROPE_SCALE environment variable is set, but will be overwritten by command line argument --rope-scale
2024-11-17 00:33:21 build: 4112 (eda7e1d4) with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
2024-11-17 00:33:21 system info: n_threads = 8, n_threads_batch = 8, total_threads = 16
2024-11-17 00:33:21
2024-11-17 00:33:21 system_info: n_threads = 8 (n_threads_batch = 8) / 16 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | AMX_INT8 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | RISCV_VECT = 0 | WASM_SIMD = 0 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 |
2024-11-17 00:33:21
2024-11-17 00:33:21 main: HTTP server is listening, hostname: 0.0.0.0, port: 8000, http threads: 15
2024-11-17 00:33:21 main: loading model
2024-11-17 00:33:22 llama_load_model_from_file: using device CUDA0 (NVIDIA GeForce RTX 3060) - 11247 MiB free
2024-11-17 00:33:22 llama_model_loader: loaded meta data with 34 key-value pairs and 339 tensors from /models/Qwen2.5-7B-Instruct-Q4_K_M.gguf (version GGUF V3 (latest))
2024-11-17 00:33:22 llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
2024-11-17 00:33:22 llama_model_loader: - kv 0: general.architecture str = qwen2
2024-11-17 00:33:22 llama_model_loader: - kv 1: general.type str = model
2024-11-17 00:33:22 llama_model_loader: - kv 2: general.name str = Qwen2.5 7B Instruct
2024-11-17 00:33:22 llama_model_loader: - kv 3: general.finetune str = Instruct
2024-11-17 00:33:22 llama_model_loader: - kv 4: general.basename str = Qwen2.5
2024-11-17 00:33:22 llama_model_loader: - kv 5: general.size_label str = 7B
2024-11-17 00:33:22 llama_model_loader: - kv 6: general.license str = apache-2.0
2024-11-17 00:33:22 llama_model_loader: - kv 7: general.license.link str = https://huggingface.co/Qwen/Qwen2.5-7...
2024-11-17 00:33:22 llama_model_loader: - kv 8: general.base_model.count u32 = 1
2024-11-17 00:33:22 llama_model_loader: - kv 9: general.base_model.0.name str = Qwen2.5 7B
2024-11-17 00:33:22 llama_model_loader: - kv 10: general.base_model.0.organization str = Qwen
2024-11-17 00:33:22 llama_model_loader: - kv 11: general.base_model.0.repo_url str = https://huggingface.co/Qwen/Qwen2.5-7B
2024-11-17 00:33:22 llama_model_loader: - kv 12: general.tags arr[str,2] = ["chat", "text-generation"]
2024-11-17 00:33:22 llama_model_loader: - kv 13: general.languages arr[str,1] = ["en"]
2024-11-17 00:33:22 llama_model_loader: - kv 14: qwen2.block_count u32 = 28
2024-11-17 00:33:22 llama_model_loader: - kv 15: qwen2.context_length u32 = 32768
2024-11-17 00:33:22 llama_model_loader: - kv 16: qwen2.embedding_length u32 = 3584
2024-11-17 00:33:22 llama_model_loader: - kv 17: qwen2.feed_forward_length u32 = 18944
2024-11-17 00:33:22 llama_model_loader: - kv 18: qwen2.attention.head_count u32 = 28
2024-11-17 00:33:22 llama_model_loader: - kv 19: qwen2.attention.head_count_kv u32 = 4
2024-11-17 00:33:22 llama_model_loader: - kv 20: qwen2.rope.freq_base f32 = 1000000.000000
2024-11-17 00:33:22 llama_model_loader: - kv 21: qwen2.attention.layer_norm_rms_epsilon f32 = 0.000001
2024-11-17 00:33:22 llama_model_loader: - kv 22: general.file_type u32 = 15
2024-11-17 00:33:22 llama_model_loader: - kv 23: tokenizer.ggml.model str = gpt2
2024-11-17 00:33:22 llama_model_loader: - kv 24: tokenizer.ggml.pre str = qwen2
2024-11-17 00:33:22 llama_model_loader: - kv 25: tokenizer.ggml.tokens arr[str,152064] = ["!", "\"", "#", "$", "%", "&", "'", ...
2024-11-17 00:33:22 llama_model_loader: - kv 26: tokenizer.ggml.token_type arr[i32,152064] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
2024-11-17 00:33:22 llama_model_loader: - kv 27: tokenizer.ggml.merges arr[str,151387] = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
2024-11-17 00:33:22 llama_model_loader: - kv 28: tokenizer.ggml.eos_token_id u32 = 151645
2024-11-17 00:33:22 llama_model_loader: - kv 29: tokenizer.ggml.padding_token_id u32 = 151643
2024-11-17 00:33:22 llama_model_loader: - kv 30: tokenizer.ggml.bos_token_id u32 = 151643
2024-11-17 00:33:22 llama_model_loader: - kv 31: tokenizer.ggml.add_bos_token bool = false
2024-11-17 00:33:22 llama_model_loader: - kv 32: tokenizer.chat_template str = {%- if tools %}\n {{- '<|im_start|>...
2024-11-17 00:33:22 llama_model_loader: - kv 33: general.quantization_version u32 = 2
2024-11-17 00:33:22 llama_model_loader: - type f32: 141 tensors
2024-11-17 00:33:22 llama_model_loader: - type q4_K: 169 tensors
2024-11-17 00:33:22 llama_model_loader: - type q6_K: 29 tensors
2024-11-17 00:33:23 llm_load_vocab: special tokens cache size = 22
2024-11-17 00:33:23 llm_load_vocab: token to piece cache size = 0.9310 MB
2024-11-17 00:33:23 llm_load_print_meta: format = GGUF V3 (latest)
2024-11-17 00:33:23 llm_load_print_meta: arch = qwen2
2024-11-17 00:33:23 llm_load_print_meta: vocab type = BPE
2024-11-17 00:33:23 llm_load_print_meta: n_vocab = 152064
2024-11-17 00:33:23 llm_load_print_meta: n_merges = 151387
2024-11-17 00:33:23 llm_load_print_meta: vocab_only = 0
2024-11-17 00:33:23 llm_load_print_meta: n_ctx_train = 32768
2024-11-17 00:33:23 llm_load_print_meta: n_embd = 3584
2024-11-17 00:33:23 llm_load_print_meta: n_layer = 28
2024-11-17 00:33:23 llm_load_print_meta: n_head = 28
2024-11-17 00:33:23 llm_load_print_meta: n_head_kv = 4
2024-11-17 00:33:23 llm_load_print_meta: n_rot = 128
2024-11-17 00:33:23 llm_load_print_meta: n_swa = 0
2024-11-17 00:33:23 llm_load_print_meta: n_embd_head_k = 128
2024-11-17 00:33:23 llm_load_print_meta: n_embd_head_v = 128
2024-11-17 00:33:23 llm_load_print_meta: n_gqa = 7
2024-11-17 00:33:23 llm_load_print_meta: n_embd_k_gqa = 512
2024-11-17 00:33:23 llm_load_print_meta: n_embd_v_gqa = 512
2024-11-17 00:33:23 llm_load_print_meta: f_norm_eps = 0.0e+00
2024-11-17 00:33:23 llm_load_print_meta: f_norm_rms_eps = 1.0e-06
2024-11-17 00:33:23 llm_load_print_meta: f_clamp_kqv = 0.0e+00
2024-11-17 00:33:23 llm_load_print_meta: f_max_alibi_bias = 0.0e+00
2024-11-17 00:33:23 llm_load_print_meta: f_logit_scale = 0.0e+00
2024-11-17 00:33:23 llm_load_print_meta: n_ff = 18944
2024-11-17 00:33:23 llm_load_print_meta: n_expert = 0
2024-11-17 00:33:23 llm_load_print_meta: n_expert_used = 0
2024-11-17 00:33:23 llm_load_print_meta: causal attn = 1
2024-11-17 00:33:23 llm_load_print_meta: pooling type = 0
2024-11-17 00:33:23 llm_load_print_meta: rope type = 2
2024-11-17 00:33:23 llm_load_print_meta: rope scaling = linear
2024-11-17 00:33:23 llm_load_print_meta: freq_base_train = 1000000.0
2024-11-17 00:33:23 llm_load_print_meta: freq_scale_train = 1
2024-11-17 00:33:23 llm_load_print_meta: n_ctx_orig_yarn = 32768
2024-11-17 00:33:23 llm_load_print_meta: rope_finetuned = unknown
2024-11-17 00:33:23 llm_load_print_meta: ssm_d_conv = 0
2024-11-17 00:33:23 llm_load_print_meta: ssm_d_inner = 0
2024-11-17 00:33:23 llm_load_print_meta: ssm_d_state = 0
2024-11-17 00:33:23 llm_load_print_meta: ssm_dt_rank = 0
2024-11-17 00:33:23 llm_load_print_meta: ssm_dt_b_c_rms = 0
2024-11-17 00:33:23 llm_load_print_meta: model type = 7B
2024-11-17 00:33:23 llm_load_print_meta: model ftype = Q4_K - Medium
2024-11-17 00:33:23 llm_load_print_meta: model params = 7.62 B
2024-11-17 00:33:23 llm_load_print_meta: model size = 4.36 GiB (4.91 BPW)
2024-11-17 00:33:23 llm_load_print_meta: general.name = Qwen2.5 7B Instruct
2024-11-17 00:33:23 llm_load_print_meta: BOS token = 151643 '<|endoftext|>'
2024-11-17 00:33:23 llm_load_print_meta: EOS token = 151645 '<|im_end|>'
2024-11-17 00:33:23 llm_load_print_meta: EOT token = 151645 '<|im_end|>'
2024-11-17 00:33:23 llm_load_print_meta: PAD token = 151643 '<|endoftext|>'
2024-11-17 00:33:23 llm_load_print_meta: LF token = 148848 'ÄĬ'
2024-11-17 00:33:23 llm_load_print_meta: FIM PRE token = 151659 '<|fim_prefix|>'
2024-11-17 00:33:23 llm_load_print_meta: FIM SUF token = 151661 '<|fim_suffix|>'
2024-11-17 00:33:23 llm_load_print_meta: FIM MID token = 151660 '<|fim_middle|>'
2024-11-17 00:33:23 llm_load_print_meta: FIM PAD token = 151662 '<|fim_pad|>'
2024-11-17 00:33:23 llm_load_print_meta: FIM REP token = 151663 '<|repo_name|>'
2024-11-17 00:33:23 llm_load_print_meta: FIM SEP token = 151664 '<|file_sep|>'
2024-11-17 00:33:23 llm_load_print_meta: EOG token = 151643 '<|endoftext|>'
2024-11-17 00:33:23 llm_load_print_meta: EOG token = 151645 '<|im_end|>'
2024-11-17 00:33:23 llm_load_print_meta: EOG token = 151662 '<|fim_pad|>'
2024-11-17 00:33:23 llm_load_print_meta: EOG token = 151663 '<|repo_name|>'
2024-11-17 00:33:23 llm_load_print_meta: EOG token = 151664 '<|file_sep|>'
2024-11-17 00:33:23 llm_load_print_meta: max token length = 256
2024-11-17 00:33:23 llm_load_tensors: offloading 28 repeating layers to GPU
2024-11-17 00:33:23 llm_load_tensors: offloading output layer to GPU
2024-11-17 00:33:23 llm_load_tensors: offloaded 29/29 layers to GPU
2024-11-17 00:33:23 llm_load_tensors: CUDA0 model buffer size = 4168.09 MiB
2024-11-17 00:33:23 llm_load_tensors: CPU model buffer size = 292.36 MiB
2024-11-17 00:34:09 .................................................................................
2024-11-17 00:34:09 llama_new_context_with_model: n_seq_max = 1
2024-11-17 00:34:09 llama_new_context_with_model: n_ctx = 131072
2024-11-17 00:34:09 llama_new_context_with_model: n_ctx_per_seq = 131072
2024-11-17 00:34:09 llama_new_context_with_model: n_batch = 2048
2024-11-17 00:34:09 llama_new_context_with_model: n_ubatch = 512
2024-11-17 00:34:09 llama_new_context_with_model: flash_attn = 1
2024-11-17 00:34:09 llama_new_context_with_model: freq_base = 1000000.0
2024-11-17 00:34:09 llama_new_context_with_model: freq_scale = 0.25
2024-11-17 00:34:09 llama_new_context_with_model: n_ctx_pre_seq (131072) > n_ctx_train (32768) -- possible training context overflow
2024-11-17 00:34:09 llama_kv_cache_init: CUDA0 KV buffer size = 7168.00 MiB
2024-11-17 00:34:09 llama_new_context_with_model: KV self size = 7168.00 MiB, K (f16): 3584.00 MiB, V (f16): 3584.00 MiB
2024-11-17 00:34:09 llama_new_context_with_model: CUDA_Host output buffer size = 0.58 MiB
2024-11-17 00:34:10 llama_new_context_with_model: CUDA0 compute buffer size = 412.00 MiB
2024-11-17 00:34:10 llama_new_context_with_model: CUDA_Host compute buffer size = 263.01 MiB
2024-11-17 00:34:10 llama_new_context_with_model: graph nodes = 875
2024-11-17 00:34:10 llama_new_context_with_model: graph splits = 2
2024-11-17 00:34:10 common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
What happened?
I'm running llama.cpp server with docker. My docker compose command set as the following:
Expectation: the rope scaling type should set to yarn and the rope scale should set to 4
Current result: the rope scaling type is still linear and the scale is still 2, which are all the default options
P.S. I even set the environment variables, and the log showed:
But somehow things still didn't work
Name and Version
build: 4112 (eda7e1d) with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
What operating system are you seeing the problem on?
Linux, Windows
Relevant log output