Eval bug: LLAMA_SET_ROWS=1 gibberish output with Dual GPU offload

[askmyteapot]

Name and Version

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: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
Device 1: Tesla P40, compute capability 6.1, VMM: no
version: 5944 (36c1532)
built with MSVC 19.44.35208.0 for x64

Operating systems

Windows

GGML backends

CUDA

Hardware

Ryzen 5800x
3090 + P40

Models

magnum-v4-22b-Q6_K.gguf
TheSkullery_L3.3-Unnamed-Exp-70B-v0.8-IQ4_XS.gguf

Problem description & steps to reproduce

Running with environment variable 'LLAMA_SET_ROWS=0' results in normal output. Setting to 1 results in gibberish

Helpful AI
21 July 2025 7:42 AM

How can I help?
12t
User
21 July 2025 7:43 AM

Can you recite for me the intro to sesame street
6.4s
85t
Helpful AI
21 July 2025 8:25 PM

Sunnynynynyy dayyy day,,
day!…
AItSunIt''''ssIts
unnme of
time for for to to
talk you play
play S S
S
ThisSThis is is is is the is a the way street best w song sest
way of way you a
to explore to be come in a
to play learn
play

However, if i restrict to just a single GPU (P40 or 3090), llama_set_rows=1 works with no issues.

First Bad Commit

Havent tested earlier versions.

Relevant log output

With LLAMA_SET_ROWS=1
Gibberish:
.\llama-server.exe -m D:\text-generation-webui\models\TheSkullery_L3.3-Unnamed-Exp-70B-v0.8-IQ4_XS.gguf -ngl 99 -ts 40/43 -fa -c 32768
.\llama-server.exe -m D:\text-generation-webui\models\magnum-v4-22b-Q6_K.gguf -ngl 99 -ts 40/43 -fa -c 32768

Sane (i cant fit 32k ctx on one GPU:
.\llama-server.exe -m D:\text-generation-webui\models\magnum-v4-22b-Q6_K.gguf -ngl 99 -fa -c 16368
.\llama-server.exe -m D:\text-generation-webui\models\magnum-v4-22b-Q6_K.gguf -ngl 53 -fa -c 32768 (CPU has 4 layers offloaded)

[ggerganov]

If you remove FA does it work?

[askmyteapot]

Still borked without -fa

[skruse]

I sporadically have the same problem on a single L40S. Strange: It recovers mid-generation and then returns to gibberish again. The slot that is used continues to have this problem, but after reload or on different slot it can't be triggered by the same prompt.

[ggerganov]

@askmyteapot Can you confirm that it also happens using llama-cli instead of llama-server?

[Panchovix]

Not OP here, but I can't replicate on Mistral 123B at least.

I have multiple GPUs, and running a Q4_K_M 123B model fully on GPU seems to work fine.

export CUDA_VISIBLE_DEVICES=3,0,1,4
pancho@fedora:/run/media/pancho/60A2FCEDA2FCC894/ChatIAs/llama.cpp/lenux/bin$ ./llama-server -m '/run/media/pancho/DE1652041651DDD9/HuggingFaceModelDownloader/Storage/GGUFs/Mistral-Large-Instruct-2411-Q4_K_M-merged.gguf' -ngl 999 -c 32768 -fa
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 RTX A6000, compute capability 8.6, VMM: yes
  Device 1: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
  Device 2: NVIDIA GeForce RTX 5090, compute capability 12.0, VMM: yes
  Device 3: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
build: 5953 (cd465d82) with cc (GCC) 14.3.1 20250523 (Red Hat 14.3.1-1) for x86_64-redhat-linux
system info: n_threads = 8, n_threads_batch = 8, total_threads = 16

system_info: n_threads = 8 (n_threads_batch = 8) / 16 | CUDA : ARCHS = 860,890,1200 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | FA_ALL_QUANTS = 1 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 

main: binding port with default address family
main: HTTP server is listening, hostname: 127.0.0.1, port: 8080, http threads: 15
main: loading model
srv    load_model: loading model '/run/media/pancho/DE1652041651DDD9/HuggingFaceModelDownloader/Storage/GGUFs/Mistral-Large-Instruct-2411-Q4_K_M-merged.gguf'
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA RTX A6000) - 48281 MiB free
llama_model_load_from_file_impl: using device CUDA1 (NVIDIA GeForce RTX 4090) - 23698 MiB free
llama_model_load_from_file_impl: using device CUDA2 (NVIDIA GeForce RTX 5090) - 29073 MiB free
llama_model_load_from_file_impl: using device CUDA3 (NVIDIA GeForce RTX 3090) - 23871 MiB free
llama_model_loader: loaded meta data with 44 key-value pairs and 795 tensors from /run/media/pancho/DE1652041651DDD9/HuggingFaceModelDownloader/Storage/GGUFs/Mistral-Large-Instruct-2411-Q4_K_M-merged.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.name str              = Mistral Large Instruct 2411
llama_model_loader: - kv   3:                            general.version str              = 2411
llama_model_loader: - kv   4:                           general.finetune str              = Instruct
llama_model_loader: - kv   5:                           general.basename str              = Mistral
llama_model_loader: - kv   6:                         general.size_label str              = Large
llama_model_loader: - kv   7:                            general.license str              = other
llama_model_loader: - kv   8:                       general.license.name str              = mrl
llama_model_loader: - kv   9:                       general.license.link str              = https://mistral.ai/licenses/MRL-0.1.md
llama_model_loader: - kv  10:                          general.languages arr[str,10]      = ["en", "fr", "de", "es", "it", "pt", ...
llama_model_loader: - kv  11:                          llama.block_count u32              = 88
llama_model_loader: - kv  12:                       llama.context_length u32              = 131072
llama_model_loader: - kv  13:                     llama.embedding_length u32              = 12288
llama_model_loader: - kv  14:                  llama.feed_forward_length u32              = 28672
llama_model_loader: - kv  15:                 llama.attention.head_count u32              = 96
llama_model_loader: - kv  16:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv  17:                       llama.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  18:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  19:                 llama.attention.key_length u32              = 128
llama_model_loader: - kv  20:               llama.attention.value_length u32              = 128
llama_model_loader: - kv  21:                          general.file_type u32              = 15
llama_model_loader: - kv  22:                           llama.vocab_size u32              = 32768
llama_model_loader: - kv  23:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv  24:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  25:                         tokenizer.ggml.pre str              = default
llama_model_loader: - kv  26:                      tokenizer.ggml.tokens arr[str,32768]   = ["<unk>", "<s>", "</s>", "[INST]", "[...
llama_model_loader: - kv  27:                      tokenizer.ggml.scores arr[f32,32768]   = [-1000.000000, -1000.000000, -1000.00...
llama_model_loader: - kv  28:                  tokenizer.ggml.token_type arr[i32,32768]   = [3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, ...
llama_model_loader: - kv  29:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  30:                tokenizer.ggml.eos_token_id u32              = 2
llama_model_loader: - kv  31:            tokenizer.ggml.unknown_token_id u32              = 0
llama_model_loader: - kv  32:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  33:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  34:                    tokenizer.chat_template str              = {{ bos_token }}{% for message in mess...
llama_model_loader: - kv  35:            tokenizer.ggml.add_space_prefix bool             = false
llama_model_loader: - kv  36:               general.quantization_version u32              = 2
llama_model_loader: - kv  37:                      quantize.imatrix.file str              = /models_out/Mistral-Large-Instruct-24...
llama_model_loader: - kv  38:                   quantize.imatrix.dataset str              = /training_dir/calibration_datav3.txt
llama_model_loader: - kv  39:             quantize.imatrix.entries_count i32              = 616
llama_model_loader: - kv  40:              quantize.imatrix.chunks_count i32              = 148
llama_model_loader: - kv  41:                                   split.no u16              = 0
llama_model_loader: - kv  42:                        split.tensors.count i32              = 795
llama_model_loader: - kv  43:                                split.count u16              = 0
llama_model_loader: - type  f32:  177 tensors
llama_model_loader: - type q4_K:  529 tensors
llama_model_loader: - type q6_K:   89 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 68.19 GiB (4.78 BPW) 
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
load: special tokens cache size = 771
load: token to piece cache size = 0.1732 MB
print_info: arch             = llama
print_info: vocab_only       = 0
print_info: n_ctx_train      = 131072
print_info: n_embd           = 12288
print_info: n_layer          = 88
print_info: n_head           = 96
print_info: n_head_kv        = 8
print_info: n_rot            = 128
print_info: n_swa            = 0
print_info: is_swa_any       = 0
print_info: n_embd_head_k    = 128
print_info: n_embd_head_v    = 128
print_info: n_gqa            = 12
print_info: n_embd_k_gqa     = 1024
print_info: n_embd_v_gqa     = 1024
print_info: f_norm_eps       = 0.0e+00
print_info: f_norm_rms_eps   = 1.0e-05
print_info: f_clamp_kqv      = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
print_info: f_logit_scale    = 0.0e+00
print_info: f_attn_scale     = 0.0e+00
print_info: n_ff             = 28672
print_info: n_expert         = 0
print_info: n_expert_used    = 0
print_info: causal attn      = 1
print_info: pooling type     = 0
print_info: rope type        = 0
print_info: rope scaling     = linear
print_info: freq_base_train  = 1000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 131072
print_info: rope_finetuned   = unknown
print_info: model type       = ?B
print_info: model params     = 122.61 B
print_info: general.name     = Mistral Large Instruct 2411
print_info: vocab type       = SPM
print_info: n_vocab          = 32768
print_info: n_merges         = 0
print_info: BOS token        = 1 '<s>'
print_info: EOS token        = 2 '</s>'
print_info: UNK token        = 0 '<unk>'
print_info: LF token         = 781 '<0x0A>'
print_info: EOG token        = 2 '</s>'
print_info: max token length = 48
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 88 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 89/89 layers to GPU
load_tensors:        CUDA0 model buffer size = 27695.44 MiB
load_tensors:        CUDA1 model buffer size = 13072.69 MiB
load_tensors:        CUDA2 model buffer size = 15479.91 MiB
load_tensors:        CUDA3 model buffer size = 13362.89 MiB
load_tensors:   CPU_Mapped model buffer size =   216.00 MiB
....................................................................................................
llama_context: constructing llama_context
llama_context: non-unified KV cache requires ggml_set_rows() - forcing unified KV cache
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 32768
llama_context: n_ctx_per_seq = 32768
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 1
llama_context: kv_unified    = true
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (32768) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_context:  CUDA_Host  output buffer size =     0.12 MiB
llama_kv_cache_unified:      CUDA0 KV buffer size =  4480.00 MiB
llama_kv_cache_unified:      CUDA1 KV buffer size =  2176.00 MiB
llama_kv_cache_unified:      CUDA2 KV buffer size =  2560.00 MiB
llama_kv_cache_unified:      CUDA3 KV buffer size =  2048.00 MiB
llama_kv_cache_unified: size = 11264.00 MiB ( 32768 cells,  88 layers,  1/ 1 seqs), K (f16): 5632.00 MiB, V (f16): 5632.00 MiB
llama_kv_cache_unified: LLAMA_SET_ROWS=0, using old ggml_cpy() method for backwards compatibility
llama_context: pipeline parallelism enabled (n_copies=4)
llama_context:      CUDA0 compute buffer size =   600.01 MiB
llama_context:      CUDA1 compute buffer size =   464.01 MiB
llama_context:      CUDA2 compute buffer size =   464.01 MiB
llama_context:      CUDA3 compute buffer size =   464.02 MiB
llama_context:  CUDA_Host compute buffer size =   280.02 MiB
llama_context: graph nodes  = 2911
llama_context: graph splits = 5
common_init_from_params: added </s> logit bias = -inf
common_init_from_params: setting dry_penalty_last_n to ctx_size = 32768
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
srv          init: initializing slots, n_slots = 1
slot         init: id  0 | task -1 | new slot n_ctx_slot = 32768
main: model loaded
main: chat template, chat_template: {{ bos_token }}{% for message in messages %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + '[/INST]' }}{% elif message['role'] == 'system' %}{{ '[SYSTEM_PROMPT] ' + message['content'] + '[/SYSTEM_PROMPT]' }}{% elif message['role'] == 'assistant' %}{{ ' ' + message['content'] + eos_token }}{% else %}{{ raise_exception('Only user, system and assistant roles are supported!') }}{% endif %}{% endfor %}, example_format: '[SYSTEM_PROMPT] You are a helpful assistant[/SYSTEM_PROMPT][INST] Hello[/INST] Hi there</s>[INST] How are you?[/INST]'
main: server is listening on http://127.0.0.1:8080 - starting the main loop
srv  update_slots: all slots are idle
srv  log_server_r: request: GET / 127.0.0.1 200
srv  log_server_r: request: GET /favicon.ico 127.0.0.1 404
srv  log_server_r: request: GET /props 127.0.0.1 200
srv  params_from_: Chat format: Content-only
slot launch_slot_: id  0 | task 0 | processing task
slot update_slots: id  0 | task 0 | new prompt, n_ctx_slot = 32768, n_keep = 0, n_prompt_tokens = 5
slot update_slots: id  0 | task 0 | kv cache rm [0, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 5, n_tokens = 5, progress = 1.000000
slot update_slots: id  0 | task 0 | prompt done, n_past = 5, n_tokens = 5
slot      release: id  0 | task 0 | stop processing: n_past = 28, truncated = 0
slot print_timing: id  0 | task 0 | 
prompt eval time =     144.52 ms /     5 tokens (   28.90 ms per token,    34.60 tokens per second)
       eval time =    1918.58 ms /    24 tokens (   79.94 ms per token,    12.51 tokens per second)
      total time =    2063.10 ms /    29 tokens
srv  update_slots: all slots are idle
srv  log_server_r: request: POST /v1/chat/completions 127.0.0.1 200
srv  params_from_: Chat format: Content-only
slot launch_slot_: id  0 | task 25 | processing task
slot update_slots: id  0 | task 25 | new prompt, n_ctx_slot = 32768, n_keep = 0, n_prompt_tokens = 42
slot update_slots: id  0 | task 25 | kv cache rm [28, end)
slot update_slots: id  0 | task 25 | prompt processing progress, n_past = 42, n_tokens = 14, progress = 0.333333
slot update_slots: id  0 | task 25 | prompt done, n_past = 42, n_tokens = 14
slot      release: id  0 | task 25 | stop processing: n_past = 205, truncated = 0
slot print_timing: id  0 | task 25 | 
prompt eval time =     174.01 ms /    14 tokens (   12.43 ms per token,    80.45 tokens per second)
       eval time =   13574.18 ms /   164 tokens (   82.77 ms per token,    12.08 tokens per second)
      total time =   13748.19 ms /   178 tokens
srv  update_slots: all slots are idle
srv  log_server_r: request: POST /v1/chat/completions 127.0.0.1 200

I compiled as

cmake -B lenux \
  -DGGML_CUDA=ON \
  -DGGML_CUDA_FA_ALL_QUANTS=ON \
  -DGGML_BLAS=OFF \
  -DCMAKE_CUDA_ARCHITECTURES="86;89;120" \
  -DGGML_MAX_CONTEXTS=2048 \
  -DGGML_LLAMA_SET_ROWS=1 \

[skruse]

I tried the compile parameters from above and it still happened for me. It seems to be triggered by parallel generations.

My setup: 6 parallel slots, and it occurs during simultaneous generation across multiple slots.

[ggerganov]

@Panchovix LLAMA_SET_ROWS is an environment variable, not a CMake variable. Read #14285 for more info.

[Panchovix]

@ggerganov nice, thanks for the headup! Tried it as a env variable but still works. Now I'm not doing parallel slots, so maybe it is related to that.

[askmyteapot]

@askmyteapot Can you confirm that it also happens using llama-cli instead of llama-server?

broken with llama-cli too

PS D:\llamacpp(built)> $Env:LLAMA_SET_ROWS=1
PS D:\llamacpp(built)> .\llama-cli.exe -m D:\text-generation-webui\models\magnum-v4-22b-Q6_K.gguf -ngl 99 -ts 40/43 -fa -c 32768
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: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
  Device 1: Tesla P40, compute capability 6.1, VMM: no
build: 5944 (36c15324) with MSVC 19.44.35208.0 for x64
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 3090) - 23306 MiB free
llama_model_load_from_file_impl: using device CUDA1 (Tesla P40) - 24319 MiB free
llama_model_loader: loaded meta data with 34 key-value pairs and 507 tensors from D:\text-generation-webui\models\magnum-v4-22b-Q6_K.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.name str              = Anthracite Core_Magnum v4 22b R4
llama_model_loader: - kv   3:                           general.finetune str              = r4
llama_model_loader: - kv   4:                           general.basename str              = anthracite-core_magnum-v4
llama_model_loader: - kv   5:                         general.size_label str              = 22B
llama_model_loader: - kv   6:                               general.tags arr[str,2]       = ["axolotl", "generated_from_trainer"]
llama_model_loader: - kv   7:                          llama.block_count u32              = 56
llama_model_loader: - kv   8:                       llama.context_length u32              = 32768
llama_model_loader: - kv   9:                     llama.embedding_length u32              = 6144
llama_model_loader: - kv  10:                  llama.feed_forward_length u32              = 16384
llama_model_loader: - kv  11:                 llama.attention.head_count u32              = 48
llama_model_loader: - kv  12:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv  13:                       llama.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  14:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  15:                 llama.attention.key_length u32              = 128
llama_model_loader: - kv  16:               llama.attention.value_length u32              = 128
llama_model_loader: - kv  17:                          general.file_type u32              = 18
llama_model_loader: - kv  18:                           llama.vocab_size u32              = 32768
llama_model_loader: - kv  19:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv  20:            tokenizer.ggml.add_space_prefix bool             = false
llama_model_loader: - kv  21:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  22:                         tokenizer.ggml.pre str              = default
llama_model_loader: - kv  23:                      tokenizer.ggml.tokens arr[str,32768]   = ["<unk>", "<s>", "</s>", "[INST]", "[...
llama_model_loader: - kv  24:                      tokenizer.ggml.scores arr[f32,32768]   = [-1000.000000, -1000.000000, -1000.00...
llama_model_loader: - kv  25:                  tokenizer.ggml.token_type arr[i32,32768]   = [3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, ...
llama_model_loader: - kv  26:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  27:                tokenizer.ggml.eos_token_id u32              = 2
llama_model_loader: - kv  28:            tokenizer.ggml.unknown_token_id u32              = 0
llama_model_loader: - kv  29:            tokenizer.ggml.padding_token_id u32              = 2
llama_model_loader: - kv  30:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  31:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  32:                    tokenizer.chat_template str              = {%- if messages[0]["role"] == "system...
llama_model_loader: - kv  33:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:  113 tensors
llama_model_loader: - type q6_K:  394 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q6_K
print_info: file size   = 17.00 GiB (6.56 BPW)
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
load: special tokens cache size = 771
load: token to piece cache size = 0.1732 MB
print_info: arch             = llama
print_info: vocab_only       = 0
print_info: n_ctx_train      = 32768
print_info: n_embd           = 6144
print_info: n_layer          = 56
print_info: n_head           = 48
print_info: n_head_kv        = 8
print_info: n_rot            = 128
print_info: n_swa            = 0
print_info: is_swa_any       = 0
print_info: n_embd_head_k    = 128
print_info: n_embd_head_v    = 128
print_info: n_gqa            = 6
print_info: n_embd_k_gqa     = 1024
print_info: n_embd_v_gqa     = 1024
print_info: f_norm_eps       = 0.0e+00
print_info: f_norm_rms_eps   = 1.0e-05
print_info: f_clamp_kqv      = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
print_info: f_logit_scale    = 0.0e+00
print_info: f_attn_scale     = 0.0e+00
print_info: n_ff             = 16384
print_info: n_expert         = 0
print_info: n_expert_used    = 0
print_info: causal attn      = 1
print_info: pooling type     = 0
print_info: rope type        = 0
print_info: rope scaling     = linear
print_info: freq_base_train  = 1000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 32768
print_info: rope_finetuned   = unknown
print_info: model type       = ?B
print_info: model params     = 22.25 B
print_info: general.name     = Anthracite Core_Magnum v4 22b R4
print_info: vocab type       = SPM
print_info: n_vocab          = 32768
print_info: n_merges         = 0
print_info: BOS token        = 1 '<s>'
print_info: EOS token        = 2 '</s>'
print_info: UNK token        = 0 '<unk>'
print_info: PAD token        = 2 '</s>'
print_info: LF token         = 781 '<0x0A>'
print_info: EOG token        = 2 '</s>'
print_info: max token length = 48
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 56 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 57/57 layers to GPU
load_tensors:        CUDA0 model buffer size =  8545.69 MiB
load_tensors:        CUDA1 model buffer size =  8703.21 MiB
load_tensors:   CPU_Mapped model buffer size =   157.50 MiB
....................................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 32768
llama_context: n_ctx_per_seq = 32768
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 1
llama_context: kv_unified    = false
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context:  CUDA_Host  output buffer size =     0.12 MiB
llama_kv_cache_unified:      CUDA0 KV buffer size =  3584.00 MiB
llama_kv_cache_unified:      CUDA1 KV buffer size =  3584.00 MiB
llama_kv_cache_unified: size = 7168.00 MiB ( 32768 cells,  56 layers,  1/ 1 seqs), K (f16): 3584.00 MiB, V (f16): 3584.00 MiB
llama_context: pipeline parallelism enabled (n_copies=4)
llama_context:      CUDA0 compute buffer size =   468.04 MiB
llama_context:      CUDA1 compute buffer size =   308.05 MiB
llama_context:  CUDA_Host compute buffer size =   268.05 MiB
llama_context: graph nodes  = 1743
llama_context: graph splits = 3
common_init_from_params: added </s> logit bias = -inf
common_init_from_params: setting dry_penalty_last_n to ctx_size = 32768
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
main: llama threadpool init, n_threads = 8
main: chat template is available, enabling conversation mode (disable it with -no-cnv)
main: chat template example:
[INST] You are a helpful assistant

Hello[/INST] Hi there</s>[INST] How are you?[/INST]

system_info: n_threads = 8 (n_threads_batch = 8) / 16 | CUDA : ARCHS = 610,860 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |

main: interactive mode on.
sampler seed: 2471128836
sampler params:
        repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
        dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 32768
        top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
        mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-n-sigma -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist
generate: n_ctx = 32768, n_batch = 2048, n_predict = -1, n_keep = 1

== Running in interactive mode. ==
 - Press Ctrl+C to interject at any time.
 - Press Return to return control to the AI.
 - To return control without starting a new line, end your input with '/'.
 - If you want to submit another line, end your input with '\'.
 - Not using system message. To change it, set a different value via -sys PROMPT


> What is the first few lines to the intro song of sesame street.
 Sunnynynyny street dayah day
,





ItItCanA-a-
CanItACan you you you

A



SunSunnynyny








SunCan









Sun

Sunny
 A
SunSun
Sun


>

[ggerganov]

Does it work with this patch:

diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index 6eb344736..12982c08f 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -708,7 +708,7 @@ llm_graph_result * llama_context::process_ubatch(const llama_ubatch & ubatch, ll
     // in order to correctly reuse a graph, it's full topology has to be uniquely determined by these parameters
     const auto gparams = graph_params(res, ubatch, mctx, gtype);
 
-    if (res->can_reuse(gparams)) {
+    if (false) {
         //LLAMA_LOG_DEBUG("%s: reusing previous graph\n", __func__);
 
         n_reused++;

[askmyteapot]

diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index 6eb344736..12982c08f 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -708,7 +708,7 @@ llm_graph_result * llama_context::process_ubatch(const llama_ubatch & ubatch, ll
// in order to correctly reuse a graph, it's full topology has to be uniquely determined by these parameters
const auto gparams = graph_params(res, ubatch, mctx, gtype);

• if (res->can_reuse(gparams)) {

• if (false) {
  //LLAMA_LOG_DEBUG("%s: reusing previous graph\n", func);

  n_reused++;
  

@ggerganov Yes. The patch fixed the gibberish output.

[thad0ctor]

Does it work with this patch:

diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index 6eb344736..12982c08f 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -708,7 +708,7 @@ llm_graph_result * llama_context::process_ubatch(const llama_ubatch & ubatch, ll
// in order to correctly reuse a graph, it's full topology has to be uniquely determined by these parameters
const auto gparams = graph_params(res, ubatch, mctx, gtype);

• if (res->can_reuse(gparams)) {

• if (false) {
  //LLAMA_LOG_DEBUG("%s: reusing previous graph\n", func);

  n_reused++;
  

The patch appears to stop the gibberish but seems to be defaulting back to unified even when LLAMA_SET_ROWS=1 as an environmental variable:

llama_context: CUDA_Host output buffer size = 0.58 MiB
llama_kv_cache_unified: CUDA0 KV buffer size = 1920.00 MiB
llama_kv_cache_unified: CUDA1 KV buffer size = 1920.00 MiB
llama_kv_cache_unified: CUDA2 KV buffer size = 1840.00 MiB
llama_kv_cache_unified: CUDA3 KV buffer size = 1840.00 MiB
llama_kv_cache_unified: size = 7520.00 MiB (131072 cells, 94 layers, 1/ 1 seqs), K (q4_1): 3760.00 MiB, V (q4_1): 3760.00 MiB

Before the patch I was getting 60 t/s of gibberish, after the patch I get 44 t/s coherent, if I turn off the LLAMA_SET_ROWS=0 or use --kv-unified I still get the same 44 t/s

EDIT: it looks like the llama_kv_cache_unified debugs are used regardless if ggml_set_rows() is used when LLAMA_SET_ROWS=1, however the patch to disable the graphs impacted performance significantly (understandably). I think an older build I was getting close to 50+ t/s so presumably it may have been working before. I'll have to look at older commits and see if anything caused this.

[ggerganov]

@thad0ctor When you have only one sequence, you always get unified KV cache, regardless of LLAMA_SET_ROWS. The regression investigated here is enabling graph reuse (#14482) when pipeline parallelism is also enabled - there seems to be a bug in this case. I'll take a look a bit later.