Deepseek-based model throws std::out_of_range exception on load

[brittlewis12]

Model: OpenCodeInterpreter-DS-6.7B (GGUFs)

This is a deepseek coder instruct-based model, llama arch, but maybe there's something distinct for it that requires special-handling?

Or maybe I did something wrong in converting these files from the original safetensors (used the same build, b2249, for converting, quantizing, and running).

Both -ngl=999 & -ngl=0 produce the same exception:

libc++abi: terminating due to uncaught exception of type std::out_of_range: unordered_map::at: key not found

llama.cpp build info

• b2249 (rev: 15499eb94227401bdc8875da6eb85c15d37068f7)
• compiled with LLAMA_METAL=1
• macOS M1 Pro

lldb stacktrace

Process 25487 stopped
* thread #1, queue = 'com.apple.main-thread', stop reason = breakpoint 1.1
    frame #0: 0x0000000188223330 libc++abi.dylib`__cxa_throw
libc++abi.dylib`__cxa_throw:
->  0x188223330 <+0>:  pacibsp
    0x188223334 <+4>:  stp    x22, x21, [sp, #-0x30]!
    0x188223338 <+8>:  stp    x20, x19, [sp, #0x10]
    0x18822333c <+12>: stp    x29, x30, [sp, #0x20]
(lldb) bt
* thread #1, queue = 'com.apple.main-thread', stop reason = breakpoint 1.1
  * frame #0: 0x0000000188223330 libc++abi.dylib`__cxa_throw
    frame #1: 0x00000001000684c0 main`std::__1::__throw_out_of_range[abi:v160006](char const*) + 60
    frame #2: 0x000000010006a790 main`llama_byte_to_token(llama_vocab const&, unsigned char) + 472
    frame #3: 0x000000010003d270 main`llama_model_load(std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> const&, llama_model&, llama_model_params&) + 1968
    frame #4: 0x000000010003ca08 main`llama_load_model_from_file + 420
    frame #5: 0x00000001000a208c main`llama_init_from_gpt_params(gpt_params&) + 96
    frame #6: 0x00000001000ed73c main`main + 2404
    frame #7: 0x0000000187ee90e0 dyld`start + 2360

full lldb output from `./main`:

(lldb) target create "./main"
Current executable set to '/Users/tito/code/llama.cpp/main' (arm64).
(lldb) settings set -- target.run-args  "-m" "/Users/tito/code/autogguf/OpenCodeInterpreter-DS-6.7B/opencodeinterpreter-ds-6.7b.Q4_K_M.gguf" "-t" "7" "--color" "--ctx_size" "4096" "--keep" "4" "--in-prefix" "<|User|>\\n" "--in-suffix" "\\n<|Assistant|>\\n" "-r" "<|User|>" "-r" "<|Assistant|>" "-r" "<|EOT|>" "-ins" "-b" "512" "-n" "-1" "--temp" "0.7" "--repeat_penalty" "1.1" "-ngl" "0"
(lldb) breakpoint set -E C++
Breakpoint 1: no locations (pending).
(lldb) run
Process 25487 launched: '/Users/tito/code/llama.cpp/main' (arm64)
2 locations added to breakpoint 1
Log start
main: build = 2249 (15499eb9)
main: built with Apple clang version 15.0.0 (clang-1500.1.0.2.5) for arm64-apple-darwin23.3.0
main: seed  = 1708707124
llama_model_loader: loaded meta data with 23 key-value pairs and 291 tensors from /Users/tito/code/autogguf/OpenCodeInterpreter-DS-6.7B/opencodeinterpreter-ds-6.7b.Q4_K_M.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.context_length u32              = 16384
llama_model_loader: - kv   3:                     llama.embedding_length u32              = 4096
llama_model_loader: - kv   4:                          llama.block_count u32              = 32
llama_model_loader: - kv   5:                  llama.feed_forward_length u32              = 11008
llama_model_loader: - kv   6:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv   7:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv   8:              llama.attention.head_count_kv u32              = 32
llama_model_loader: - kv   9:     llama.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  10:                       llama.rope.freq_base f32              = 100000.000000
llama_model_loader: - kv  11:                    llama.rope.scaling.type str              = linear
llama_model_loader: - kv  12:                  llama.rope.scaling.factor f32              = 4.000000
llama_model_loader: - kv  13:                          general.file_type u32              = 15
llama_model_loader: - kv  14:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,32256]   = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  16:                      tokenizer.ggml.scores arr[f32,32256]   = [-1000.000000, -1000.000000, -1000.00...
llama_model_loader: - kv  17:                  tokenizer.ggml.token_type arr[i32,32256]   = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 32013
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 32021
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 32014
llama_model_loader: - kv  21:                    tokenizer.chat_template str              = {%- set found_item = false -%}\n{%- fo...
llama_model_loader: - kv  22:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   65 tensors
llama_model_loader: - type q4_K:  193 tensors
llama_model_loader: - type q6_K:   33 tensors
Process 25487 stopped
* thread #1, queue = 'com.apple.main-thread', stop reason = breakpoint 1.1
    frame #0: 0x0000000188223330 libc++abi.dylib`__cxa_throw
libc++abi.dylib`__cxa_throw:
->  0x188223330 <+0>:  pacibsp
    0x188223334 <+4>:  stp    x22, x21, [sp, #-0x30]!
    0x188223338 <+8>:  stp    x20, x19, [sp, #0x10]
    0x18822333c <+12>: stp    x29, x30, [sp, #0x20]
(lldb) bt
* thread #1, queue = 'com.apple.main-thread', stop reason = breakpoint 1.1
  * frame #0: 0x0000000188223330 libc++abi.dylib`__cxa_throw
    frame #1: 0x00000001000684c0 main`std::__1::__throw_out_of_range[abi:v160006](char const*) + 60
    frame #2: 0x000000010006a790 main`llama_byte_to_token(llama_vocab const&, unsigned char) + 472
    frame #3: 0x000000010003d270 main`llama_model_load(std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> const&, llama_model&, llama_model_params&) + 1968
    frame #4: 0x000000010003ca08 main`llama_load_model_from_file + 420
    frame #5: 0x00000001000a208c main`llama_init_from_gpt_params(gpt_params&) + 96
    frame #6: 0x00000001000ed73c main`main + 2404
    frame #7: 0x0000000187ee90e0 dyld`start + 2360

conversion info

$ python3.11 ./convert.py OpenCodeInterpreter-DS-6.7B \
  --outtype f16 \
  --outfile opencodeinterpreter-ds-6.7b.fp16.gguf \
  --vocab-type hfft \
  --pad-vocab

[brittlewis12]

whoops, it was indeed my mistake in the conversion!

turns out that, while the base instruct model uses a fast tokenizer, this model instead uses the regular llama tokenizer. which means I should've converted with BPE!

reconverted & quantized and what do you know, it runs great.

---

I doubt it's worth investigating the crash on its own given the incorrectly produced model file.

But maybe there could be a way to detect this sort of mistake at conversion time to short circuit this process? Auto vocab-type detection would be beneficial, but that's out of the scope of this issue.

[ggerganov]

reconverted & quantized and what do you know, it runs great.

Huh, that's surprising. There is a long pending PR that I thought needs to be merged to support DeepSeek models: #5464. It should fix some tokenization problems AFAICT + add conversion

I'm surprised that it worked for you

[brittlewis12]

the updated fp16 conversion and quants just finished uploading: hf link

it does seem to work fine tho! I haven't tested it too extensively, but:

full output running main:

❯ ./opencodeinterp.sh
Log start
main: build = 2249 (15499eb9)
main: built with Apple clang version 15.0.0 (clang-1500.1.0.2.5) for arm64-apple-darwin23.3.0
main: seed  = 1708713048
llama_model_loader: loaded meta data with 24 key-value pairs and 291 tensors from /Users/tito/code/autogguf/OpenCodeInterpreter-DS-6.7B/opencodeinterpreter-ds-6.7b.Q4_K_M.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.context_length u32              = 16384
llama_model_loader: - kv   3:                     llama.embedding_length u32              = 4096
llama_model_loader: - kv   4:                          llama.block_count u32              = 32
llama_model_loader: - kv   5:                  llama.feed_forward_length u32              = 11008
llama_model_loader: - kv   6:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv   7:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv   8:              llama.attention.head_count_kv u32              = 32
llama_model_loader: - kv   9:     llama.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  10:                       llama.rope.freq_base f32              = 100000.000000
llama_model_loader: - kv  11:                    llama.rope.scaling.type str              = linear
llama_model_loader: - kv  12:                  llama.rope.scaling.factor f32              = 4.000000
llama_model_loader: - kv  13:                          general.file_type u32              = 15
llama_model_loader: - kv  14:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,32256]   = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  16:                      tokenizer.ggml.scores arr[f32,32256]   = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv  17:                  tokenizer.ggml.token_type arr[i32,32256]   = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  18:                      tokenizer.ggml.merges arr[str,31757]   = ["Ġ Ġ", "Ġ t", "Ġ a", "i n", "h e...
llama_model_loader: - kv  19:                tokenizer.ggml.bos_token_id u32              = 32013
llama_model_loader: - kv  20:                tokenizer.ggml.eos_token_id u32              = 32021
llama_model_loader: - kv  21:            tokenizer.ggml.padding_token_id u32              = 32014
llama_model_loader: - kv  22:                    tokenizer.chat_template str              = {%- set found_item = false -%}\n{%- fo...
llama_model_loader: - kv  23:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   65 tensors
llama_model_loader: - type q4_K:  193 tensors
llama_model_loader: - type q6_K:   33 tensors
llm_load_vocab: mismatch in special tokens definition ( 243/32256 vs 256/32256 ).
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          = 32256
llm_load_print_meta: n_merges         = 31757
llm_load_print_meta: n_ctx_train      = 16384
llm_load_print_meta: n_embd           = 4096
llm_load_print_meta: n_head           = 32
llm_load_print_meta: n_head_kv        = 32
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            = 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-06
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: n_ff             = 11008
llm_load_print_meta: n_expert         = 0
llm_load_print_meta: n_expert_used    = 0
llm_load_print_meta: rope scaling     = linear
llm_load_print_meta: freq_base_train  = 100000.0
llm_load_print_meta: freq_scale_train = 0.25
llm_load_print_meta: n_yarn_orig_ctx  = 16384
llm_load_print_meta: rope_finetuned   = unknown
llm_load_print_meta: model type       = 7B
llm_load_print_meta: model ftype      = Q4_K - Medium
llm_load_print_meta: model params     = 6.74 B
llm_load_print_meta: model size       = 3.80 GiB (4.84 BPW)
llm_load_print_meta: general.name     = .
llm_load_print_meta: BOS token        = 32013 '<｜begin▁of▁sentence｜>'
llm_load_print_meta: EOS token        = 32021 '<|EOT|>'
llm_load_print_meta: PAD token        = 32014 '<｜end▁of▁sentence｜>'
llm_load_print_meta: LF token         = 126 'Ä'
llm_load_tensors: ggml ctx size =    0.22 MiB
ggml_backend_metal_buffer_from_ptr: allocated buffer, size =  3821.77 MiB, ( 3821.83 / 21845.34)
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:      Metal buffer size =  3821.76 MiB
llm_load_tensors:        CPU buffer size =    70.88 MiB
..................................................................................................
llama_new_context_with_model: n_ctx      = 4096
llama_new_context_with_model: freq_base  = 100000.0
llama_new_context_with_model: freq_scale = 0.25
ggml_metal_init: allocating
ggml_metal_init: found device: Apple M1 Pro
ggml_metal_init: picking default device: Apple M1 Pro
ggml_metal_init: default.metallib not found, loading from source
ggml_metal_init: GGML_METAL_PATH_RESOURCES = nil
ggml_metal_init: loading '/Users/tito/code/llama.cpp/ggml-metal.metal'
ggml_metal_init: GPU name:   Apple M1 Pro
ggml_metal_init: GPU family: MTLGPUFamilyApple7  (1007)
ggml_metal_init: GPU family: MTLGPUFamilyCommon3 (3003)
ggml_metal_init: GPU family: MTLGPUFamilyMetal3  (5001)
ggml_metal_init: simdgroup reduction support   = true
ggml_metal_init: simdgroup matrix mul. support = true
ggml_metal_init: hasUnifiedMemory              = true
ggml_metal_init: recommendedMaxWorkingSetSize  = 22906.50 MB
ggml_backend_metal_buffer_type_alloc_buffer: allocated buffer, size =  2048.00 MiB, ( 5871.64 / 21845.34)
llama_kv_cache_init:      Metal KV buffer size =  2048.00 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:        CPU input buffer size   =    17.04 MiB
ggml_backend_metal_buffer_type_alloc_buffer: allocated buffer, size =   296.03 MiB, ( 6167.67 / 21845.34)
llama_new_context_with_model:      Metal compute buffer size =   296.02 MiB
llama_new_context_with_model:        CPU compute buffer size =     8.00 MiB
llama_new_context_with_model: graph splits (measure): 3

system_info: n_threads = 7 / 10 | AVX = 0 | AVX_VNNI = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | SSSE3 = 0 | VSX = 0 | MATMUL_INT8 = 0 |
main: interactive mode on.
Reverse prompt: '<|User|>'
Reverse prompt: '<|Assistant|>'
Reverse prompt: '<|EOT|>'
Reverse prompt: '### Instruction:

'
Input prefix: '<|User|>\n'
Input suffix: '\n<|Assistant|>\n'
sampling:
	repeat_last_n = 64, repeat_penalty = 1.100, 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 = 0.700
	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 = 4096, n_batch = 512, n_predict = -1, n_keep = 1


== Running in interactive mode. ==
 - Press Ctrl+C to interject at any time.
 - Press Return to return control to LLaMa.
 - To return control without starting a new line, end your input with '/'.
 - If you want to submit another line, end your input with '\'.


> <|User|>\nwrite fibonacci with a cache hashmap in ruby\n<|Assistant|>\ndef fibonacci(n, cache = {})
  if n == 0 || n == 1
    return n
  end

  unless cache[n]
     cache[n] = fibonacci(n - 1, cache) + fibonacci(n - 2, cache)
  end

  cache[n]
end

# Example usage:
puts fibonacci(5) # Output: 3

This function uses a technique called memoization to store the result of expensive function calls and reusing them when same inputs occur again. This greatly improves performance for recursive algorithms like this one, by reducing redundant calculations. The cache hashmap serves as our "memory" in which we store previously calculated results of fibonacci(n - 1) and fibonacci(n - 2).

The base case is if n equals to 0 or 1, return n itself because the first two numbers in Fibonacci sequence are 0 and 1. If not, it checks whether the result of this function call has been calculated before. If yes, it returns that result directly from the cache hashmap; otherwise, it calculates it by calling fibonacci(n - 1) + fibonacci(n - 2), stores it in the cache hashmap for future use and then return it.

> <|User|>\n

llama_print_timings:        load time =    5624.60 ms
llama_print_timings:      sample time =     188.29 ms /   305 runs   (    0.62 ms per token,  1619.81 tokens per second)
llama_print_timings: prompt eval time =     499.83 ms /    36 tokens (   13.88 ms per token,    72.02 tokens per second)
llama_print_timings:        eval time =   12152.07 ms /   305 runs   (   39.84 ms per token,    25.10 tokens per second)
llama_print_timings:       total time =  187992.14 ms /   341 tokens

that script just calls main with in-prefix/-suffix, ngl, temp, etc.