search

ggerganov / llama.cpp

LLM inference in C/C++
MIT License
67.89k stars 9.73k forks source link

The server was not compiled for multimodal or the model projector can't be loaded #8595

Closed simon-krannig closed 2 months ago

simon-krannig commented 3 months ago

What happened?

Hey everyone, i am currently trying to set up llama cpp server with a llava vision model. When using the llama-llava-cli, everything works just fine:

./llama-llava-cli -m ../llava-v1.6-vicuna-7b/Llava-V1.6-Vicuna-7B-F32.gguf --mmproj vit/mmproj-model-f16.gguf --image /root/Gfp-wisconsin-madison-the-nature-boardwalk.jpg -c 4096 -ngl 15000 -p "The image depicts a place. Describe this place. Would my dog, a natural explorer, be happy there?" --temp 0
(...)
encode_image_with_clip: 5 segments encoded in   320.06 ms
encode_image_with_clip: image embedding created: 2880 tokens

encode_image_with_clip: image encoded in   433.94 ms by CLIP (    0.15 ms per image patch)

 The image shows a serene and picturesque landscape. It appears to be a rural area with a long, narrow pathway leading through a field of tall grass. The pathway is flanked by a grassy area on one side and a line of trees or bushes on the other, providing a natural boundary. The sky is partly cloudy, suggesting a pleasant day with some sunshine.

Your dog, as a natural explorer, would likely be happy in such a setting. The open space and the opportunity to explore the grassy areas and the pathway would be appealing to a dog that enjoys running and sniffing around. The presence of trees and bushes would also offer shade and additional exploration opportunities. However, it's important to ensure that your dog is well-behaved and under control, especially if there are other animals or people around.

However, i cant get the same functionality to work when using llama.cpp server. Running this command ...

./llama-server -m ../llava-v1.6-vicuna-7b/Llava-V1.6-Vicuna-7B-F32.gguf --mmproj vit/mmproj-model-f16.gguf --image /root/Gfp-wisconsin-madison-the-nature-boardwalk.jpg -c 4096 -ngl 15000 --host 0.0.0.0 --port 8010

... and trying to use the image with the llama.cpp server frontend results in the error message:

The server was not compiled for multimodal or the model projector can't be loaded

Trying to use the completion or openAI style chat api with images in payload results in the image simply being ignored entirely (base64 or url, doesn't matter) The server logs also show no indication of CLIP being used. Does anybody have an idea what might be going on? This reddit thread from months ago shows this functionality in action.

Name and Version

./llama-llava-cli --version version: 3417 (3d0e4367) built with cc (Ubuntu 13.2.0-23ubuntu4) 13.2.0 for x86_64-linux-gnu

./llama-server --version version: 3417 (3d0e4367) built with cc (Ubuntu 13.2.0-23ubuntu4) 13.2.0 for x86_64-linux-gnu

What operating system are you seeing the problem on?

Linux

Relevant log output

./llama-server -m ../llava-v1.6-vicuna-7b/Llava-V1.6-Vicuna-7B-F32.gguf --mmproj vit/mmproj-model-f16.gguf --image /root/Gfp-wisconsin-madison-the-nature-boardwalk.jpg -c 4096 -ngl 15000 --host 0.0.0.0 --port 8010
INFO [                    main] build info | tid="131625621454848" timestamp=1721418528 build=3417 commit="3d0e4367"
INFO [                    main] system info | tid="131625621454848" timestamp=1721418528 n_threads=4 n_threads_batch=-1 total_threads=4 system_info="AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 1 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 | "
llama_model_loader: loaded meta data with 25 key-value pairs and 291 tensors from ../llava-v1.6-vicuna-7b/Llava-V1.6-Vicuna-7B-F32.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              = Llava v1.6 Vicuna 7b
llama_model_loader: - kv   2:                           general.basename str              = llava-v1.6-vicuna
llama_model_loader: - kv   3:                         general.size_label str              = 7.1B
llama_model_loader: - kv   4:                               general.tags arr[str,1]       = ["image-text-to-text"]
llama_model_loader: - kv   5:                           llama.vocab_size u32              = 32000
llama_model_loader: - kv   6:                       llama.context_length u32              = 4096
llama_model_loader: - kv   7:                     llama.embedding_length u32              = 4096
llama_model_loader: - kv   8:                          llama.block_count u32              = 32
llama_model_loader: - kv   9:                  llama.feed_forward_length u32              = 11008
llama_model_loader: - kv  10:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv  11:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv  12:              llama.attention.head_count_kv u32              = 32
llama_model_loader: - kv  13:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  14:                       llama.rope.freq_base f32              = 10000.000000
llama_model_loader: - kv  15:                          general.file_type u32              = 0
llama_model_loader: - kv  16:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  17:                      tokenizer.ggml.tokens arr[str,32000]   = ["<unk>", "<s>", "</s>", "<0x00>", "<...
llama_model_loader: - kv  18:                      tokenizer.ggml.scores arr[f32,32000]   = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv  19:                  tokenizer.ggml.token_type arr[i32,32000]   = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...
llama_model_loader: - kv  20:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  21:                tokenizer.ggml.eos_token_id u32              = 2
llama_model_loader: - kv  22:            tokenizer.ggml.padding_token_id u32              = 0
llama_model_loader: - kv  23:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  24:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - type  f32:  291 tensors
llm_load_vocab: special tokens cache size = 3
llm_load_vocab: token to piece cache size = 0.1684 MB
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = llama
llm_load_print_meta: vocab type       = SPM
llm_load_print_meta: n_vocab          = 32000
llm_load_print_meta: n_merges         = 0
llm_load_print_meta: vocab_only       = 0
llm_load_print_meta: n_ctx_train      = 4096
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  = 10000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn  = 4096
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       = 7B
llm_load_print_meta: model ftype      = all F32
llm_load_print_meta: model params     = 6.74 B
llm_load_print_meta: model size       = 25.10 GiB (32.00 BPW) 
llm_load_print_meta: general.name     = Llava v1.6 Vicuna 7b
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: PAD token        = 0 '<unk>'
llm_load_print_meta: LF token         = 13 '<0x0A>'
llm_load_print_meta: max token length = 48
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 2 CUDA devices:
  Device 0: Quadro RTX 8000, compute capability 7.5, VMM: yes
  Device 1: Quadro RTX 8000, compute capability 7.5, VMM: yes
llm_load_tensors: ggml ctx size =    0.41 MiB
llm_load_tensors: offloading 32 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 33/33 layers to GPU
llm_load_tensors:        CPU buffer size =   500.00 MiB
llm_load_tensors:      CUDA0 buffer size = 13124.53 MiB
llm_load_tensors:      CUDA1 buffer size = 12080.48 MiB
....................................................................................................
llama_new_context_with_model: n_ctx      = 4096
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  = 10000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:      CUDA0 KV buffer size =  1088.00 MiB
llama_kv_cache_init:      CUDA1 KV buffer size =   960.00 MiB
llama_new_context_with_model: KV self size  = 2048.00 MiB, K (f16): 1024.00 MiB, V (f16): 1024.00 MiB
llama_new_context_with_model:  CUDA_Host  output buffer size =     0.24 MiB
llama_new_context_with_model: pipeline parallelism enabled (n_copies=4)
llama_new_context_with_model:      CUDA0 compute buffer size =   352.01 MiB
llama_new_context_with_model:      CUDA1 compute buffer size =   352.02 MiB
llama_new_context_with_model:  CUDA_Host compute buffer size =    40.02 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 3
INFO [                    init] initializing slots | tid="131625621454848" timestamp=1721418536 n_slots=1
INFO [                    init] new slot | tid="131625621454848" timestamp=1721418536 id_slot=0 n_ctx_slot=4096
INFO [                    main] model loaded | tid="131625621454848" timestamp=1721418536
INFO [                    main] chat template | tid="131625621454848" timestamp=1721418536 chat_example="<|im_start|>system\nYou are a helpful assistant<|im_end|>\n<|im_start|>user\nHello<|im_end|>\n<|im_start|>assistant\nHi there<|im_end|>\n<|im_start|>user\nHow are you?<|im_end|>\n<|im_start|>assistant\n" built_in=true
INFO [                    main] HTTP server listening | tid="131625621454848" timestamp=1721418536 n_threads_http="3" port="8010" hostname="0.0.0.0"
INFO [            update_slots] all slots are idle | tid="131625621454848" timestamp=1721418536
INFO [      log_server_request] request | tid="131624169635840" timestamp=1721418628 remote_addr="XXX" remote_port=54157 status=200 method="GET" path="/" params={}
INFO [      log_server_request] request | tid="131624169635840" timestamp=1721418628 remote_addr="XXX" remote_port=54157 status=200 method="GET" path="/index.js" params={}
INFO [      log_server_request] request | tid="131624159150080" timestamp=1721418628 remote_addr="XXX" remote_port=54158 status=200 method="GET" path="/completion.js" params={}
INFO [      log_server_request] request | tid="131624148664320" timestamp=1721418628 remote_addr="XXX" remote_port=54159 status=200 method="GET" path="/json-schema-to-grammar.mjs" params={}
INFO [   launch_slot_with_task] slot is processing task | tid="131625621454848" timestamp=1721418637 id_slot=0 id_task=0
INFO [            update_slots] kv cache rm [p0, end) | tid="131625621454848" timestamp=1721418637 id_slot=0 id_task=0 p0=0
INFO [      log_server_request] request | tid="131624159150080" timestamp=1721418638 remote_addr="10.243.231.194" remote_port=54160 status=200 method="POST" path="/completion" params={}
INFO [            update_slots] slot released | tid="131625621454848" timestamp=1721418638 id_slot=0 id_task=0 n_ctx=4096 n_past=69 n_system_tokens=0 n_cache_tokens=69 truncated=false
INFO [            update_slots] all slots are idle | tid="131625621454848" timestamp=1721418638
simon-krannig commented 3 months ago

Just realized that multimodal support got dropped (temporarily) for llama-server.

Any updates on when this feature will be available again?

ZhenyaPav commented 3 months ago

Would also love to know

joshuaipwork commented 3 months ago

Do we know what commit or PR support got dropped?

ali0une commented 3 months ago

Latest release to support LLaVA with server was B2356

ElhamAhmedian commented 3 months ago

So I can not use moondream2 or Bunny VLM's with the server? What is the alternative?

ali0une commented 3 months ago

LMStudio

create a directory LMStudio/models/WHATEVER/local/moondream2/ copy moondream2-mmproj-f16.gguf and moondream2-text-model-f16.gguf into it. Select Alpaca for preset.

Profit.

ElhamAhmedian commented 3 months ago

@ali0une Thanks, but I need this as a server, could that work?

jabberjabberjabber commented 3 months ago

@ali0une Thanks, but I need this as a server, could that work?

KoboldCPP maintains LLava support.

github-actions[bot] commented 2 months ago

This issue was closed because it has been inactive for 14 days since being marked as stale.