add 128k yarn context for Qwen

[robbiemu]

see: discussion

@bartowski1182 -- can I ask you to try this if you have a 7b+ Qwen2.5 handy ? I dont mind testing it but I thought it would be nice if a 3rd party did it.

quick instructions (correct me if Im wrong):

add rope scaling like:

  "rope_scaling": {
    "factor": 4.0,
    "original_max_position_embeddings": 32768,
    "type": "yarn"
  }

change max_position_embeddings to factor * orig_mpe:

  "max_position_embeddings": 131072,  

[robbiemu]

Thanks to @ggerganov , I think I've verified my change working.

./convert_hf_to_gguf.py --outtype bf16 --outfile /Users/Shared/Public/huggingface/Qwen2.5-Coder-7B-Instruct/Qwen2.5_7b_bf16.gguf /Users/Shared/Public/huggingface/Qwen2.5-Coder-7B-Instruct
INFO:hf-to-gguf:Loading model: Qwen2.5-Coder-7B-Instruct
INFO:gguf.gguf_writer:gguf: This GGUF file is for Little Endian only
INFO:hf-to-gguf:Exporting model...
...
INFO:hf-to-gguf:Set model quantization version
INFO:gguf.gguf_writer:Writing the following files:
INFO:gguf.gguf_writer:/Users/Shared/Public/huggingface/Qwen2.5-Coder-7B-Instruct/Qwen2.5_7b_bf16.gguf: n_tensors = 339, total_size = 15.2G
huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...
To disable this warning, you can either:
	- Avoid using `tokenizers` before the fork if possible
	- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)
Writing: 100%|██████████████████████████████████████| 15.2G/15.2G [00:44<00:00, 339Mbyte/s]
INFO:hf-to-gguf:Model successfully exported to /Users/Shared/Public/huggingface/Qwen2.5-Coder-7B-Instruct/Qwen2.5_7b_bf16.gguf

and then

llama-passkey -m ./Qwen2.5_7b_bf16_2.gguf --junk 5420
build: 4176 (9a4b79bc) with Apple clang version 16.0.0 (clang-1600.0.26.4) for arm64-apple-darwin24.1.0
llama_load_model_from_file: using device Metal (Apple M3 Max) - 40959 MiB free
llama_model_loader: loaded meta data with 38 key-value pairs and 339 tensors from ./Qwen2.5_7b_bf16_2.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              = qwen2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Qwen2.5 Coder 7B Instruct
llama_model_loader: - kv   3:                           general.finetune str              = Instruct
llama_model_loader: - kv   4:                           general.basename str              = Qwen2.5-Coder
llama_model_loader: - kv   5:                         general.size_label str              = 7B
llama_model_loader: - kv   6:                            general.license str              = apache-2.0
llama_model_loader: - kv   7:                       general.license.link str              = https://huggingface.co/Qwen/Qwen2.5-C...
llama_model_loader: - kv   8:                   general.base_model.count u32              = 1
llama_model_loader: - kv   9:                  general.base_model.0.name str              = Qwen2.5 Coder 7B
llama_model_loader: - kv  10:          general.base_model.0.organization str              = Qwen
llama_model_loader: - kv  11:              general.base_model.0.repo_url str              = https://huggingface.co/Qwen/Qwen2.5-C...
llama_model_loader: - kv  12:                               general.tags arr[str,6]       = ["code", "codeqwen", "chat", "qwen", ...
llama_model_loader: - kv  13:                          general.languages arr[str,1]       = ["en"]
llama_model_loader: - kv  14:                          qwen2.block_count u32              = 28
llama_model_loader: - kv  15:                       qwen2.context_length u32              = 131072
llama_model_loader: - kv  16:                     qwen2.embedding_length u32              = 3584
llama_model_loader: - kv  17:                  qwen2.feed_forward_length u32              = 18944
llama_model_loader: - kv  18:                 qwen2.attention.head_count u32              = 28
llama_model_loader: - kv  19:              qwen2.attention.head_count_kv u32              = 4
llama_model_loader: - kv  20:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  21:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  22:                          general.file_type u32              = 32
llama_model_loader: - kv  23:                    qwen2.rope.scaling.type str              = yarn
llama_model_loader: - kv  24:                  qwen2.rope.scaling.factor f32              = 4.000000
llama_model_loader: - kv  25: qwen2.rope.scaling.original_context_length u32              = 32768
llama_model_loader: - kv  26:     qwen2.rope.scaling.yarn_log_multiplier f32              = 0.100000
llama_model_loader: - kv  27:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  28:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  29:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  30:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  31:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  32:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  33:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  34:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  35:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  36:                    tokenizer.chat_template str              = {%- if tools %}\n    {{- '<|im_start|>...
llama_model_loader: - kv  37:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:  141 tensors
llama_model_loader: - type bf16:  198 tensors
llm_load_vocab: special tokens cache size = 22
llm_load_vocab: token to piece cache size = 0.9310 MB
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = qwen2
llm_load_print_meta: vocab type       = BPE
llm_load_print_meta: n_vocab          = 152064
llm_load_print_meta: n_merges         = 151387
llm_load_print_meta: vocab_only       = 0
llm_load_print_meta: n_ctx_train      = 131072
llm_load_print_meta: n_embd           = 3584
llm_load_print_meta: n_layer          = 28
llm_load_print_meta: n_head           = 28
llm_load_print_meta: n_head_kv        = 4
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            = 7
llm_load_print_meta: n_embd_k_gqa     = 512
llm_load_print_meta: n_embd_v_gqa     = 512
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: f_logit_scale    = 0.0e+00
llm_load_print_meta: n_ff             = 18944
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        = 2
llm_load_print_meta: rope scaling     = yarn
llm_load_print_meta: freq_base_train  = 1000000.0
llm_load_print_meta: freq_scale_train = 0.25
llm_load_print_meta: n_ctx_orig_yarn  = 32768
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: ssm_dt_b_c_rms   = 0
llm_load_print_meta: model type       = 7B
llm_load_print_meta: model ftype      = BF16
llm_load_print_meta: model params     = 7.62 B
llm_load_print_meta: model size       = 14.19 GiB (16.00 BPW) 
llm_load_print_meta: general.name     = Qwen2.5 Coder 7B Instruct
llm_load_print_meta: BOS token        = 151643 '<|endoftext|>'
llm_load_print_meta: EOS token        = 151645 '<|im_end|>'
llm_load_print_meta: EOT token        = 151645 '<|im_end|>'
llm_load_print_meta: PAD token        = 151643 '<|endoftext|>'
llm_load_print_meta: LF token         = 148848 'ÄĬ'
llm_load_print_meta: FIM PRE token    = 151659 '<|fim_prefix|>'
llm_load_print_meta: FIM SUF token    = 151661 '<|fim_suffix|>'
llm_load_print_meta: FIM MID token    = 151660 '<|fim_middle|>'
llm_load_print_meta: FIM PAD token    = 151662 '<|fim_pad|>'
llm_load_print_meta: FIM REP token    = 151663 '<|repo_name|>'
llm_load_print_meta: FIM SEP token    = 151664 '<|file_sep|>'
llm_load_print_meta: EOG token        = 151643 '<|endoftext|>'
llm_load_print_meta: EOG token        = 151645 '<|im_end|>'
llm_load_print_meta: EOG token        = 151662 '<|fim_pad|>'
llm_load_print_meta: EOG token        = 151663 '<|repo_name|>'
llm_load_print_meta: EOG token        = 151664 '<|file_sep|>'
llm_load_print_meta: max token length = 256
llm_load_tensors: offloading 28 repeating layers to GPU
llm_load_tensors: offloading output layer to GPU
llm_load_tensors: offloaded 29/29 layers to GPU
llm_load_tensors: Metal_Mapped model buffer size = 12447.28 MiB
llm_load_tensors:   CPU_Mapped model buffer size = 14526.27 MiB
........................................................................................
llama_new_context_with_model: n_seq_max     = 1
llama_new_context_with_model: n_ctx         = 131104
llama_new_context_with_model: n_ctx_per_seq = 131104
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     = 1000000.0
llama_new_context_with_model: freq_scale    = 0.25
llama_new_context_with_model: n_ctx_pre_seq (131104) > n_ctx_train (131072) -- possible training context overflow
ggml_metal_init: allocating
ggml_metal_init: found device: Apple M3 Max
ggml_metal_init: picking default device: Apple M3 Max
ggml_metal_init: using embedded metal library
ggml_metal_init: GPU name:   Apple M3 Max
ggml_metal_init: GPU family: MTLGPUFamilyApple9  (1009)
ggml_metal_init: GPU family: MTLGPUFamilyCommon3 (3003)
ggml_metal_init: GPU family: MTLGPUFamilyMetal3  (5001)
ggml_metal_init: simdgroup reduction   = true
ggml_metal_init: simdgroup matrix mul. = true
ggml_metal_init: has bfloat            = true
ggml_metal_init: use bfloat            = false
ggml_metal_init: hasUnifiedMemory      = true
ggml_metal_init: recommendedMaxWorkingSetSize  = 42949.67 MB
ggml_metal_init: skipping kernel_get_rows_bf16                     (not supported)
ggml_metal_init: skipping kernel_mul_mv_bf16_f32                   (not supported)
ggml_metal_init: skipping kernel_mul_mv_bf16_f32_1row              (not supported)
ggml_metal_init: skipping kernel_mul_mv_bf16_f32_l4                (not supported)
ggml_metal_init: skipping kernel_mul_mv_bf16_bf16                  (not supported)
ggml_metal_init: skipping kernel_mul_mv_id_bf16_f32                (not supported)
ggml_metal_init: skipping kernel_mul_mm_bf16_f32                   (not supported)
ggml_metal_init: skipping kernel_mul_mm_id_bf16_f32                (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h64           (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h80           (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h96           (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h112          (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h128          (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h256          (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_vec_bf16_h128      (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_vec_bf16_h256      (not supported)
ggml_metal_init: skipping kernel_cpy_f32_bf16                      (not supported)
ggml_metal_init: skipping kernel_cpy_bf16_f32                      (not supported)
ggml_metal_init: skipping kernel_cpy_bf16_bf16                     (not supported)
llama_kv_cache_init:      Metal KV buffer size =  7169.75 MiB
llama_new_context_with_model: KV self size  = 7169.75 MiB, K (f16): 3584.88 MiB, V (f16): 3584.88 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.58 MiB
llama_new_context_with_model:      Metal compute buffer size =  7453.81 MiB
llama_new_context_with_model:        CPU compute buffer size =   297.00 MiB
llama_new_context_with_model: graph nodes  = 986
llama_new_context_with_model: graph splits = 395 (with bs=512), 283 (with bs=1)

main: n_len = 130161, n_ctx = 131072, n_kv_req = 131104, n_grp = 1, n_batch = 2048, n_junk = 5420, i_pos = 547

prefix tokens: 30
prompt tokens: 130145
main: processed: [     0,   2048)
main: processed: [  2048,   4096)
main: processed: [  4096,   6144)
main: processed: [  6144,   8192)
main: processed: [  8192,  10240)
main: processed: [ 10240,  12288)
main: processed: [ 12288,  14336)
main: processed: [ 14336,  16384)
main: processed: [ 16384,  18432)
main: processed: [ 18432,  20480)
main: processed: [ 20480,  22528)
main: processed: [ 22528,  24576)
main: processed: [ 24576,  26624)
main: processed: [ 26624,  28672)
main: processed: [ 28672,  30720)
main: processed: [ 30720,  32768)
main: processed: [ 32768,  34816)
main: processed: [ 34816,  36864)
main: processed: [ 36864,  38912)
main: processed: [ 38912,  40960)
main: processed: [ 40960,  43008)
main: processed: [ 43008,  45056)
main: processed: [ 45056,  47104)
main: processed: [ 47104,  49152)
main: processed: [ 49152,  51200)
main: processed: [ 51200,  53248)
main: processed: [ 53248,  55296)
main: processed: [ 55296,  57344)
main: processed: [ 57344,  59392)
main: processed: [ 59392,  61440)
main: processed: [ 61440,  63488)
main: processed: [ 63488,  65536)
main: processed: [ 65536,  67584)
main: processed: [ 67584,  69632)
main: processed: [ 69632,  71680)
main: processed: [ 71680,  73728)
main: processed: [ 73728,  75776)
main: processed: [ 75776,  77824)
main: processed: [ 77824,  79872)
main: processed: [ 79872,  81920)
main: processed: [ 81920,  83968)
main: processed: [ 83968,  86016)
main: processed: [ 86016,  88064)
main: processed: [ 88064,  90112)
main: processed: [ 90112,  92160)
main: processed: [ 92160,  94208)
main: processed: [ 94208,  96256)
main: processed: [ 96256,  98304)
main: processed: [ 98304, 100352)
main: processed: [100352, 102400)
main: processed: [102400, 104448)
main: processed: [104448, 106496)
main: processed: [106496, 108544)
main: processed: [108544, 110592)
main: processed: [110592, 112640)
main: processed: [112640, 114688)
main: processed: [114688, 116736)
main: processed: [116736, 118784)
main: processed: [118784, 120832)
main: processed: [120832, 122880)
main: processed: [122880, 124928)
main: processed: [124928, 126976)
main: processed: [126976, 129024)
main: processed: [129024, 130145)

main: passkey = 25250, inserted at position 547 / 5420 (token pos: ~13134)

 What is the pass key? The pass key is 25250. Remember it. 25250

main: decoded 16 tokens in 13.52 s, speed: 1.18 t/s

llama_perf_context_print:        load time = 1721741.78 ms
llama_perf_context_print: prompt eval time = 1720734.48 ms / 130145 tokens (   13.22 ms per token,    75.63 tokens per second)
llama_perf_context_print:        eval time =   13509.73 ms /    16 runs   (  844.36 ms per token,     1.18 tokens per second)
llama_perf_context_print:       total time = 1735260.27 ms / 130161 tokens

ggml_metal_free: deallocating

note, it is a shame that it hid it so early .. on a previous run using only 1/2 the context it was clearly after the 32k original context window:

...
main: processed: [ 59392,  60065)
...
main: passkey = 25250, inserted at position 1807 / 2500 (token pos: ~43414)

 What is the pass key? The pass key is 25250. Remember it. 25250

main: decoded 16 tokens in 11.66 s, speed: 1.37 t/s

llama_perf_context_print:        load time =  571207.64 ms
llama_perf_context_print: prompt eval time =  570533.64 ms / 60065 tokens (    9.50 ms per token,   105.28 tokens per second)
llama_perf_context_print:        eval time =   11647.58 ms /    16 runs   (  727.97 ms per token,     1.37 tokens per second)
llama_perf_context_print:       total time =  582865.27 ms / 60081 tokens

I've converted this from draft for review :)

[BodhiHu]

Hi @robbiemu , does this PR mean the 128k YaRN context will be enabled by default when running with like Qwen2.5 7B models ?
How do we verify this when running llama-cli or llama-server ?

When testing with llama-server with inputs over 10K, we did not seem to see much difference:

./build/bin/llama-server -m qwen2.5-7b-instruct-q4_k_m.gguf -t 10 -fa --host 0.0.0.0 --port 26137

Thanks a lot.

[robbiemu]

No, you need to follow the instructions the qwen team specified, adding a section to the config.json when generating your gguf, to enable the YaRN long context.