getting SLOW imatrix completion on NVIDIA

[robbiemu]

I'm trying to do a robust quantization over 34 languages, so I have a rather large dataset for the llama-imatrix file. On my m3 Max macbook, this will take >260hrs, and I've had some crashes it seems every time I wake the screensaver, so I decided to try using vast.ai hosting. I am using the ghcr.io/ggerganov/llama.cpp:full-cuda image from this repo.

Something definitely is amiss though, On my macbook m3 max, the ETA to complete from llama-imatrix is ~260hrs

The first time I tries, the system I chose was an AMD Epic with 4x4090. the ETA was 660 hrs. I was expecting ~8 fold improvement, not ~3 fold slowdown.

I used a script to execute the command though, and it didn't record the commands used, I worried it might not have captured the number of gpus correctly and ended up passing bad params. So, I tried again. This time on a Xeon server with 4x4090s.

/app/llama-imatrix -m /workspace/salamandra-7b_bf16.gguf -f /workspace/imatrix-dataset.txt -o imatrix.dat -ngl -1 --n-gpu-layers -1 -sm layer -ts 1,1,1,1 --ctx-size 8192 --rope-freq-base 10000.0 --top-p 0.95 --temp 0 --repeat-penalty 1.2
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 4 CUDA devices:
  Device 0: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
  Device 1: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
  Device 2: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
  Device 3: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
build: 3931 (73afe681) with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
llama_model_loader: loaded meta data with 29 key-value pairs and 291 tensors from /workspace/salamandra-7b_bf16.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.type str              = model
llama_model_loader: - kv   2:                         general.size_label str              = 7.8B
llama_model_loader: - kv   3:                            general.license str              = apache-2.0
llama_model_loader: - kv   4:                               general.tags arr[str,1]       = ["text-generation"]
llama_model_loader: - kv   5:                          general.languages arr[str,36]      = ["bg", "ca", "code", "cs", "cy", "da"...
llama_model_loader: - kv   6:                          llama.block_count u32              = 32
llama_model_loader: - kv   7:                       llama.context_length u32              = 8192
llama_model_loader: - kv   8:                     llama.embedding_length u32              = 4096
llama_model_loader: - kv   9:                  llama.feed_forward_length u32              = 11008
llama_model_loader: - kv  10:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv  11:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv  12:                       llama.rope.freq_base f32              = 10000.000000
llama_model_loader: - kv  13:     llama.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  14:                          general.file_type u32              = 32
llama_model_loader: - kv  15:                           llama.vocab_size u32              = 256000
llama_model_loader: - kv  16:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv  17:            tokenizer.ggml.add_space_prefix bool             = true
llama_model_loader: - kv  18:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  19:                         tokenizer.ggml.pre str              = default
llama_model_loader: - kv  20:                      tokenizer.ggml.tokens arr[str,256000]  = ["<unk>", "<s>", "</s>", "<pad>", "<|...
llama_model_loader: - kv  21:                      tokenizer.ggml.scores arr[f32,256000]  = [-1000.000000, -1000.000000, -1000.00...
llama_model_loader: - kv  22:                  tokenizer.ggml.token_type arr[i32,256000]  = [3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, ...
llama_model_loader: - kv  23:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  24:                tokenizer.ggml.eos_token_id u32              = 2
llama_model_loader: - kv  25:            tokenizer.ggml.unknown_token_id u32              = 0
llama_model_loader: - kv  26:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  27:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  28:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   65 tensors
llama_model_loader: - type bf16:  226 tensors
llm_load_vocab: control token:     99 '<|reserved_token_94|>' is not marked as EOG
llm_load_vocab: control token:     79 '<|reserved_token_74|>' is not marked as EOG
llm_load_vocab: control token:     35 '<|reserved_token_30|>' is not marked as EOG
llm_load_vocab: control token:     44 '<|reserved_token_39|>' is not marked as EOG
llm_load_vocab: control token:     24 '<|reserved_token_19|>' is not marked as EOG
llm_load_vocab: control token:     29 '<|reserved_token_24|>' is not marked as EOG
llm_load_vocab: control token:     28 '<|reserved_token_23|>' is not marked as EOG
llm_load_vocab: control token:    101 '<|reserved_token_96|>' is not marked as EOG
llm_load_vocab: control token:     88 '<|reserved_token_83|>' is not marked as EOG
llm_load_vocab: control token:      2 '</s>' is not marked as EOG
llm_load_vocab: control token:     90 '<|reserved_token_85|>' is not marked as EOG
llm_load_vocab: control token:     87 '<|reserved_token_82|>' is not marked as EOG
llm_load_vocab: control token:      4 '<|im_start|>' is not marked as EOG
llm_load_vocab: control token:     86 '<|reserved_token_81|>' is not marked as EOG
llm_load_vocab: control token:     36 '<|reserved_token_31|>' is not marked as EOG
llm_load_vocab: control token:     38 '<|reserved_token_33|>' is not marked as EOG
llm_load_vocab: control token:     48 '<|reserved_token_43|>' is not marked as EOG
llm_load_vocab: control token:     89 '<|reserved_token_84|>' is not marked as EOG
llm_load_vocab: control token:      7 '<|reserved_token_2|>' is not marked as EOG
llm_load_vocab: control token:     40 '<|reserved_token_35|>' is not marked as EOG
llm_load_vocab: control token:     20 '<|reserved_token_15|>' is not marked as EOG
llm_load_vocab: control token:     91 '<|reserved_token_86|>' is not marked as EOG
llm_load_vocab: control token:      1 '<s>' is not marked as EOG
llm_load_vocab: control token:     70 '<|reserved_token_65|>' is not marked as EOG
llm_load_vocab: control token:     75 '<|reserved_token_70|>' is not marked as EOG
llm_load_vocab: control token:     12 '<|reserved_token_7|>' is not marked as EOG
llm_load_vocab: control token:     52 '<|reserved_token_47|>' is not marked as EOG
llm_load_vocab: control token:     72 '<|reserved_token_67|>' is not marked as EOG
llm_load_vocab: control token:      9 '<|reserved_token_4|>' is not marked as EOG
llm_load_vocab: control token:     74 '<|reserved_token_69|>' is not marked as EOG
llm_load_vocab: control token:     94 '<|reserved_token_89|>' is not marked as EOG
llm_load_vocab: control token:     77 '<|reserved_token_72|>' is not marked as EOG
llm_load_vocab: control token:     76 '<|reserved_token_71|>' is not marked as EOG
llm_load_vocab: control token:     41 '<|reserved_token_36|>' is not marked as EOG
llm_load_vocab: control token:     42 '<|reserved_token_37|>' is not marked as EOG
llm_load_vocab: control token:     22 '<|reserved_token_17|>' is not marked as EOG
llm_load_vocab: control token:     32 '<|reserved_token_27|>' is not marked as EOG
llm_load_vocab: control token:     18 '<|reserved_token_13|>' is not marked as EOG
llm_load_vocab: control token:     66 '<|reserved_token_61|>' is not marked as EOG
llm_load_vocab: control token:     96 '<|reserved_token_91|>' is not marked as EOG
llm_load_vocab: control token:     93 '<|reserved_token_88|>' is not marked as EOG
llm_load_vocab: control token:     85 '<|reserved_token_80|>' is not marked as EOG
llm_load_vocab: control token:     47 '<|reserved_token_42|>' is not marked as EOG
llm_load_vocab: control token:     61 '<|reserved_token_56|>' is not marked as EOG
llm_load_vocab: control token:     71 '<|reserved_token_66|>' is not marked as EOG
llm_load_vocab: control token:     51 '<|reserved_token_46|>' is not marked as EOG
llm_load_vocab: control token:     43 '<|reserved_token_38|>' is not marked as EOG
llm_load_vocab: control token:      8 '<|reserved_token_3|>' is not marked as EOG
llm_load_vocab: control token:     14 '<|reserved_token_9|>' is not marked as EOG
llm_load_vocab: control token:     39 '<|reserved_token_34|>' is not marked as EOG
llm_load_vocab: control token:     68 '<|reserved_token_63|>' is not marked as EOG
llm_load_vocab: control token:     50 '<|reserved_token_45|>' is not marked as EOG
llm_load_vocab: control token:     46 '<|reserved_token_41|>' is not marked as EOG
llm_load_vocab: control token:     80 '<|reserved_token_75|>' is not marked as EOG
llm_load_vocab: control token:     56 '<|reserved_token_51|>' is not marked as EOG
llm_load_vocab: control token:      6 '<|reserved_token_1|>' is not marked as EOG
llm_load_vocab: control token:     84 '<|reserved_token_79|>' is not marked as EOG
llm_load_vocab: control token:     16 '<|reserved_token_11|>' is not marked as EOG
llm_load_vocab: control token:     17 '<|reserved_token_12|>' is not marked as EOG
llm_load_vocab: control token:     64 '<|reserved_token_59|>' is not marked as EOG
llm_load_vocab: control token:     83 '<|reserved_token_78|>' is not marked as EOG
llm_load_vocab: control token:      3 '<pad>' is not marked as EOG
llm_load_vocab: control token:     19 '<|reserved_token_14|>' is not marked as EOG
llm_load_vocab: control token:    102 '<|reserved_token_97|>' is not marked as EOG
llm_load_vocab: control token:     81 '<|reserved_token_76|>' is not marked as EOG
llm_load_vocab: control token:      0 '<unk>' is not marked as EOG
llm_load_vocab: control token:     23 '<|reserved_token_18|>' is not marked as EOG
llm_load_vocab: control token:     98 '<|reserved_token_93|>' is not marked as EOG
llm_load_vocab: control token:     15 '<|reserved_token_10|>' is not marked as EOG
llm_load_vocab: control token:     33 '<|reserved_token_28|>' is not marked as EOG
llm_load_vocab: control token:     60 '<|reserved_token_55|>' is not marked as EOG
llm_load_vocab: control token:     65 '<|reserved_token_60|>' is not marked as EOG
llm_load_vocab: control token:     78 '<|reserved_token_73|>' is not marked as EOG
llm_load_vocab: control token:     31 '<|reserved_token_26|>' is not marked as EOG
llm_load_vocab: control token:    103 '<|reserved_token_98|>' is not marked as EOG
llm_load_vocab: control token:     55 '<|reserved_token_50|>' is not marked as EOG
llm_load_vocab: control token:     82 '<|reserved_token_77|>' is not marked as EOG
llm_load_vocab: control token:     63 '<|reserved_token_58|>' is not marked as EOG
llm_load_vocab: control token:     54 '<|reserved_token_49|>' is not marked as EOG
llm_load_vocab: control token:     26 '<|reserved_token_21|>' is not marked as EOG
llm_load_vocab: control token:     73 '<|reserved_token_68|>' is not marked as EOG
llm_load_vocab: control token:     25 '<|reserved_token_20|>' is not marked as EOG
llm_load_vocab: control token:     45 '<|reserved_token_40|>' is not marked as EOG
llm_load_vocab: control token:     30 '<|reserved_token_25|>' is not marked as EOG
llm_load_vocab: control token:     10 '<|reserved_token_5|>' is not marked as EOG
llm_load_vocab: control token:     49 '<|reserved_token_44|>' is not marked as EOG
llm_load_vocab: control token:     27 '<|reserved_token_22|>' is not marked as EOG
llm_load_vocab: control token:     67 '<|reserved_token_62|>' is not marked as EOG
llm_load_vocab: control token:     13 '<|reserved_token_8|>' is not marked as EOG
llm_load_vocab: control token:     57 '<|reserved_token_52|>' is not marked as EOG
llm_load_vocab: control token:     37 '<|reserved_token_32|>' is not marked as EOG
llm_load_vocab: control token:     59 '<|reserved_token_54|>' is not marked as EOG
llm_load_vocab: control token:     58 '<|reserved_token_53|>' is not marked as EOG
llm_load_vocab: control token:     11 '<|reserved_token_6|>' is not marked as EOG
llm_load_vocab: control token:     97 '<|reserved_token_92|>' is not marked as EOG
llm_load_vocab: control token:     69 '<|reserved_token_64|>' is not marked as EOG
llm_load_vocab: control token:     92 '<|reserved_token_87|>' is not marked as EOG
llm_load_vocab: control token:     62 '<|reserved_token_57|>' is not marked as EOG
llm_load_vocab: control token:     53 '<|reserved_token_48|>' is not marked as EOG
llm_load_vocab: control token:     95 '<|reserved_token_90|>' is not marked as EOG
llm_load_vocab: control token:     21 '<|reserved_token_16|>' is not marked as EOG
llm_load_vocab: control token:    100 '<|reserved_token_95|>' is not marked as EOG
llm_load_vocab: control token:     34 '<|reserved_token_29|>' is not marked as EOG
llm_load_vocab: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
llm_load_vocab: special tokens cache size = 104
llm_load_vocab: token to piece cache size = 1.8842 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          = 256000
llm_load_print_meta: n_merges         = 0
llm_load_print_meta: vocab_only       = 0
llm_load_print_meta: n_ctx_train      = 8192
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        = 8
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            = 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-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             = 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  = 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: ssm_dt_b_c_rms   = 0
llm_load_print_meta: model type       = 8B
llm_load_print_meta: model ftype      = BF16
llm_load_print_meta: model params     = 7.77 B
llm_load_print_meta: model size       = 14.47 GiB (16.00 BPW) 
llm_load_print_meta: general.name     = n/a
llm_load_print_meta: BOS token        = 1 '<s>'
llm_load_print_meta: EOS token        = 2 '</s>'
llm_load_print_meta: EOT token        = 5 '<|im_end|>'
llm_load_print_meta: UNK token        = 0 '<unk>'
llm_load_print_meta: LF token         = 145 '<0x0A>'
llm_load_print_meta: EOG token        = 2 '</s>'
llm_load_print_meta: EOG token        = 5 '<|im_end|>'
llm_load_print_meta: max token length = 72
llm_load_tensors: ggml ctx size =    0.14 MiB
llm_load_tensors: offloading 0 repeating layers to GPU
llm_load_tensors: offloaded 0/33 layers to GPU
llm_load_tensors:        CPU buffer size = 14817.02 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  = 10000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:  CUDA_Host KV buffer size =  1024.00 MiB
llama_new_context_with_model: KV self size  = 1024.00 MiB, K (f16):  512.00 MiB, V (f16):  512.00 MiB
llama_new_context_with_model:  CUDA_Host  output buffer size =     0.98 MiB
llama_new_context_with_model:      CUDA0 compute buffer size =    24.00 MiB
llama_new_context_with_model:  CUDA_Host compute buffer size =   552.01 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 196

system_info: n_threads = 72 (n_threads_batch = 72) / 144 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | RISCV_VECT = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 | 
compute_imatrix: tokenizing the input ..

This ETA was 636 hours, sorry I cut it off on the version of the log that I downloaded.

Why is it taking almost 3x as long to do this on a 4x4090 machine as it is locally?

[slaren]

The CUDA backend also does not support BF16, so most of the model is running on the CPU. Try a F16 model instead.

Also note that -ngl -1 does not work the way you might expect, no layers will be offloaded that way. Use a large number to offload the entire model instead, eg. -ngl 99.

[robbiemu]

ahaha when you said "keep in mind that BF16 is not supported in the Metal backend" last time I thought that it was unique to Metal :D

Damn, but I do want to do this in native bf16 since the original is. Is there no acceleration available to me?

(even still, my Mac laptop outperforming a Xeon 3x is still kinda surprising)

[slaren]

F16 is just as good in essentially every case. Most models are normalized so that the weights have a distribution of $\mathcal{N}(0,1)$, so the chance of the model having any values that cannot be represented in F16 is essentially zero. I think that the increased range of BF16 may have a higher importance during training, but during inference it is not necessary. For creating an imatrix, Q8_0 would also work well.

[robbiemu]

I .. thank you. can you give me a reference regarding that latter claim? I.. the purpose to creating these imatrices is so when I quantize it, I minimize ppl loss (over the target languages). You're saying, if I understand this correctly, I could go straight to Q8_0 with no imatrix, generate the imatrix from that, then go back and use that imatrix to recreate q8 and all the stronger quants?

[slaren]

See for example #5222. Admittedly, the results may be different with different models, but typically Q8_0 is very close to F16. Q8_0 cannot use an imatrix (it does not need it), but you can use the imatrix created with a Q8_0 model to create the smaller quants.

[robbiemu]

wow this has been such a treasure of a discussion, thank you @slaren
"Q8_0 cannot use an imatrix (it does not need it)" was another thing new to me. I was paying the imatrix to it anyway at that quantization level 🤣