search

ggerganov / llama.cpp

LLM inference in C/C++
MIT License
65.88k stars 9.46k forks source link

Bug: JSON Schema - enum behind a $ref generates an object with unrestricted properties #8073

Closed cikkle closed 1 month ago

cikkle commented 3 months ago

What happened?

I'm using the json_schema feature in llama-server. Using a simple prompt like Write a dialog between Alice and Biff, if I send a schema like:

{
    "type": "array",
    "minItems": 15,
    "maxItems": 15,
    "items": { "$ref": "#/$defs/TALK" },

    "$defs": {
        "TALK": {
            "type": "object",
            "required": [ "character", "emote", "dialog" ],
            "properties": {
                "character": { "enum": [ "Alice", "Biff"] },
                "emote": { "enum": ["EXCLAMATION", "CONFUSION", "CHEERFUL", "LOVE", "ANGRY", "NERVOUS", "ANNOYED", "SILENCE", "INSPIRED", "SLEEPING"] },
                "dialog": {
                    "type": "string",
                    "minLength": 1,
                    "maxLength": 200
                }
            }
        }
    }
}

I get back an array of responses in the format I'd expect, like:

{ "character": "Alice", "emote": "SILENCE", "dialog": "I'm just saying, it's not like you to be so... quiet. Is everything alright?" } {"character": "Biff", "emote": "NERVOUS", "dialog": "Yeah, everything's fine. Just... busy. You know how it is." }

Things stop working right if I try to put the enums in separate definitions. The following schema:

{
    "type": "array",
    "minItems": 15,
    "maxItems": 15,
    "items": { "$ref": "#/$defs/TALK" },

    "$defs": {
        "characters": { "enum": ["Biff", "Alice"] },
        "emotes": { "enum": ["EXCLAMATION", "CONFUSION", "CHEERFUL", "LOVE", "ANGRY"] },

        "TALK": {
            "type": "object",
            "required": [ "character", "emote", "dialog" ],
            "properties": {
                "character": { "$ref": "#/$defs/characters" },
                "emote": { "$ref": "#/$defs/emotes" },
                "dialog": {
                    "type": "string",
                    "minLength": 1,
                    "maxLength": 200
                }
            }
        }
    }
}

...gives me arbitrary things like:

{ "character": {"name": "Alice","description": "Alice, a young woman, has a bright and curious expression on her face."},
{"emotion": "curious"}
 { "character": {"name": "Biff","description": "Biff, a friendly-looking man, has a warm smile and a hint of mischief in his eyes."},
{"emotion": "amused"}

The output should follow the same format in both, but I get an object with random properties in place of the enum, and possibly more random things afterward (in this run, it was a bonus object tagging along, but it can vary).

Notably if I reorder the properties to put "dialog" before "character" I'll actually get the dialog property and string I asked for, so things only seem to go off the rails when it reaches one of the referenced enums.

I'm aware json_schema currently has some known bugs and features yet to implemented, but I didn't see anything in the readme I thought this would fall under. Terminal output from llama-server doesn't appear to show anything relevant but it's included for completeness.

Name and Version

o0@hades:~/ai/llama.cpp$ ./llama-cli --version version: 3203 (b5a5f34e) built 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

Relevant log output

INFO [                    main] build info | tid="139722331939776" timestamp=1719130838 build=3203 commit="b5a5f34e"
INFO [                    main] system info | tid="139722331939776" timestamp=1719130838 n_threads=12 n_threads_batch=-1 total_threads=24 system_info="AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | 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 26 key-value pairs and 291 tensors from ../models/text/L3-8B-Stheno-v3.2-Q8_0.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              = L3-8B-Stheno-v3.2
llama_model_loader: - kv   2:                          llama.block_count u32              = 32
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.feed_forward_length u32              = 14336
llama_model_loader: - kv   6:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv   7:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv   8:                       llama.rope.freq_base f32              = 500000.000000
llama_model_loader: - kv   9:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  10:                          general.file_type u32              = 7
llama_model_loader: - kv  11:                           llama.vocab_size u32              = 128256
llama_model_loader: - kv  12:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = llama-bpe
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,128256]  = ["!", "\"", "#", "$", "%", "&", "'", ...
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              = 128009
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: - kv  22:                      quantize.imatrix.file str              = /models/L3-8B-Stheno-v3.2-GGUF/L3-8B-...
llama_model_loader: - kv  23:                   quantize.imatrix.dataset str              = /training_data/calibration_datav3.txt
llama_model_loader: - kv  24:             quantize.imatrix.entries_count i32              = 224
llama_model_loader: - kv  25:              quantize.imatrix.chunks_count i32              = 125
llama_model_loader: - type  f32:   65 tensors
llama_model_loader: - type q8_0:  226 tensors
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: n_ctx_train      = 8192
llm_load_print_meta: n_embd           = 4096
llm_load_print_meta: n_head           = 32
llm_load_print_meta: n_head_kv        = 8
llm_load_print_meta: n_layer          = 32
llm_load_print_meta: n_rot            = 128
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      = Q8_0
llm_load_print_meta: model params     = 8.03 B
llm_load_print_meta: model size       = 7.95 GiB (8.50 BPW)
llm_load_print_meta: general.name     = L3-8B-Stheno-v3.2
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
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:   no
ggml_cuda_init: CUDA_USE_TENSOR_CORES: yes
ggml_cuda_init: found 2 ROCm devices:
  Device 0: Radeon RX 7900 XTX, compute capability 11.0, VMM: no
  Device 1: Radeon RX 7900 XTX, compute capability 11.0, VMM: no
llm_load_tensors: ggml ctx size =    0.44 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:      ROCm0 buffer size =  3757.53 MiB
llm_load_tensors:      ROCm1 buffer size =  3847.80 MiB
llm_load_tensors:        CPU buffer size =   532.31 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:      ROCm0 KV buffer size =   416.50 MiB
llama_kv_cache_init:      ROCm1 KV buffer size =   367.50 MiB
llama_new_context_with_model: KV self size  =  784.00 MiB, K (q8_0):  272.00 MiB, V (f16):  512.00 MiB
llama_new_context_with_model:  ROCm_Host  output buffer size =     0.98 MiB
llama_new_context_with_model: pipeline parallelism enabled (n_copies=4)
llama_new_context_with_model:      ROCm0 compute buffer size =   640.01 MiB
llama_new_context_with_model:      ROCm1 compute buffer size =   640.02 MiB
llama_new_context_with_model:  ROCm_Host compute buffer size =    72.02 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 3
INFO [                    init] initializing slots | tid="139722331939776" timestamp=1719130849 n_slots=1
INFO [                    init] new slot | tid="139722331939776" timestamp=1719130849 id_slot=0 n_ctx_slot=8192
INFO [                    main] model loaded | tid="139722331939776" timestamp=1719130849
INFO [                    main] chat template | tid="139722331939776" timestamp=1719130849 chat_example="<|start_header_id|>system<|end_header_id|>\n\nYou are a helpful assistant<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHello<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nHi there<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHow are you?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n" built_in=true
INFO [                    main] HTTP server listening | tid="139722331939776" timestamp=1719130849 n_threads_http="23" port="5000" hostname="0.0.0.0"
INFO [            update_slots] all slots are idle | tid="139722331939776" timestamp=1719130849
INFO [   launch_slot_with_task] slot is processing task | tid="139722331939776" timestamp=1719131079 id_slot=0 id_task=0
INFO [            update_slots] kv cache rm [p0, end) | tid="139722331939776" timestamp=1719131079 id_slot=0 id_task=0 p0=0
INFO [           print_timings] prompt eval time     =     111.88 ms /    55 tokens (    2.03 ms per token,   491.61 tokens per second) | tid="139722331939776" timestamp=1719131141 id_slot=0 id_task=0 t_prompt_processing=111.878 n_prompt_tokens_processed=55 t_token=2.0341454545454547 n_tokens_second=491.60692897620623
INFO [           print_timings] generation eval time =   61940.54 ms /  1522 runs   (   40.70 ms per token,    24.57 tokens per second) | tid="139722331939776" timestamp=1719131141 id_slot=0 id_task=0 t_token_generation=61940.538 n_decoded=1522 t_token=40.696805519053875 n_tokens_second=24.57195318516607
INFO [           print_timings]           total time =   62052.42 ms | tid="139722331939776" timestamp=1719131141 id_slot=0 id_task=0 t_prompt_processing=111.878 t_token_generation=61940.538 t_total=62052.416
INFO [            update_slots] slot released | tid="139722331939776" timestamp=1719131141 id_slot=0 id_task=0 n_ctx=8192 n_past=1576 n_system_tokens=0 n_cache_tokens=0 truncated=false
INFO [            update_slots] all slots are idle | tid="139722331939776" timestamp=1719131141
INFO [            update_slots] all slots are idle | tid="139722331939776" timestamp=1719131141
INFO [            update_slots] all slots are idle | tid="139722331939776" timestamp=1719131253
ochafik commented 3 months ago

Hi @cikkle, thanks for the report!

As you can see in the docs the support for external $refs hasn't been implemented in the C++ json schema -> grammar converter yet (will need to use CURL, it's on my todo list).

I guess we should probably find a way to issue warnings about it or even hard-fail (currently we just silently degrade to "anything goes" so the items in your example are just any json object), but in the meantime as a workaround you might wanna try and run python examples/json_schema_to_grammar.py schema.json and pass the resulting grammar to your server call, if possible. Or possibly, paste the schema in your prompt to rely on a mix of good will from the model your items and schema constraint only for the high-level shape of the output.

Cheers

cikkle commented 3 months ago

Sorry, I might have a pretty basic misunderstanding as far as the spec, the docs, or the terms used. I took "external" to mean a reference to a schema in another file; the characters and emotes in the $def block in the second example should be all local, right? Even given the notes in the docs I thought that would work.

Anyhow, for my use case I don't have much of a problem with just copying enums around the schema where they need to go, but the fallback behavior did surprise me, so thanks for looking at this.

ochafik commented 3 months ago

Sorry, I might have a pretty basic misunderstanding as far as the spec, the docs, or the terms used. I took "external" to mean a reference to a schema in another file; the characters and emotes in the $def block in the second example should be all local, right? Even given the notes in the docs I thought that would work.

Ohh sorry I completely misread your bug report, thanks for the clarification 🫣

Anyhow, for my use case I don't have much of a problem with just copying enums around the schema where they need to go, but the fallback behavior did surprise me, so thanks for looking at this.

It looks like it's a bug specific to the C++ implementation of the JSON Schema -> Grammar conversion, I'll try to send a fix shortly.

In the meantime you can use python examples/json_schema_to_grammar.py schema.json which seems to work with your example.

github-actions[bot] commented 1 month ago

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