Eval bug: DeepSeek-R1-UD-Q2_K_XL output broken

[joesixpaq]

Name and Version

I experience gibberish with DeepSeek-R1-UD-Q2_K_XL by unsloth (checked with SHA256)

In my case, this gibberish output started with e1e8e09.

I eventually managed to isolate the latest still working commit: 6f67cf1

The most recent tested commit which is still not working is 9f2da58

Operating systems

Linux

GGML backends

CUDA

Hardware

1x RTX 3090, Intel Xeon E5-2640 v3, 1TB RAM

Models

DeepSeek-R1-UD-Q2_K_XL by unsloth

Problem description & steps to reproduce

Senseless output with partially Chinese characters

First Bad Commit

e1e8e09

Relevant log output

#!/bin/bash

if [ -t 0 ]; then
    CPU0="--physcpubind=16,17,18,19,20,21,22,23 --membind=0"
    CPU1="--physcpubind=8,10,12,14,24,26,28,30 --membind=1"

    declare -a MODEL_ALIASES=(
        "DeepSeek R1 Q2_K_XL"
        "Qwen3-32B-UD-Q4_K_XL"
    )
    
    declare -a MODEL_PATHS=(
        "/mnt/AI/LLM/DeepSeek-R1-UD-Q2_K_XL/DeepSeek-R1-UD-Q2_K_XL-00001-of-00005.gguf"
        "/mnt/AI/LLM/Qwen3-32B-UD-Q4_K_XL/Qwen3-32B-UD-Q4_K_XL.gguf"
    )

    # CPU Selection
    echo "Select CPU configuration:"
    select cpu_opt in "CPU0" "CPU1"; do
        case $cpu_opt in
            "CPU0") SELECTED_CPU="$CPU0"; break;;
            "CPU1") SELECTED_CPU="$CPU1"; break;;
            *) echo "Invalid option, please choose 1 or 2";;
        esac
    done
    
    # Model Selection
    echo "Select model:"
    select model_alias in "${MODEL_ALIASES[@]}"; do
        if [[ -n "$model_alias" ]] && (( REPLY >= 1 && REPLY <= ${#MODEL_ALIASES[@]} )); then
            MODEL_PATH="${MODEL_PATHS[$((REPLY-1))]}"
            break
        else
            echo "Invalid selection. Please enter a number between 1 and ${#MODEL_ALIASES[@]}."
        fi
    done
    
    read -p "Server port [8000]: " PORT
    PORT=${PORT:-8000}
    read -p "Context size [16384]: " CTX
    CTX=${CTX:-16384}
    read -p "GPU layers [0]: " NGL
    NGL=${NGL:-0}

    SERVER_PARAMS="
    --port $PORT
    --model $MODEL_PATH
    --n-gpu-layers $NGL
    --ctx-size $CTX
    --no-mmap
    --temp 0.6
    --cache-type-k q4_0
    --threads 8
    --predict 16384
    --host 0.0.0.0
    --batch-size 4096
    --device CUDA0
    "
    # Start server with line buffering
    
    # COMMIT="e1e8e0991ffd9e99a445c6812bb519d5bac9f4b5" # FIRST BROKEN
    # COMMIT="6f67cf1f480926391ad75ff746e0a021214bf70c" # LAST GOOD
    
    # COMMIT="8afbd968182909cf93fb15959fc867b6dd3adb53" # GOOD newest 04.05.2025 for Qwen3 but not DS-R1

    COMMIT="9f2da5871f4bbd205b8a3b952cdc76283218d595" 

    PATH2APP="$HOME/workspace/LLAMA_CPP/$COMMIT/llama.cpp/build/bin/llama-server"
    echo $PATH2APP
    numactl $SELECTED_CPU stdbuf -oL "$PATH2APP" $SERVER_PARAMS > server.log 2>&1 &
    
    SERVER_PID=$!

    # Monitor log for startup completion
    echo "Waiting for server to start..."
    (
        tail -f server.log | while IFS= read -r line; do
            # Print line to terminal
            echo "$line"
            # Check for magic string
            if [[ "$line" == *"starting the main loop"* ]]; then
                pkill -P $$ tail  # Kill tail process
                exit 0
            fi
        done
    ) & TAIL_PID=$!

    # Wait for monitoring process
    wait $TAIL_PID 2>/dev/null

    # Open browser if successful
    if [ $? -eq 0 ]; then
        echo "Server ready! Launching browser..."
        xdg-open "http://localhost:$PORT" 2>/dev/null
    else
        echo "ERROR: Server failed to start"
        kill $SERVER_PID 2>/dev/null
        exit 1
    fi

    # Cleanup
    wait $SERVER_PID
    rm server.log

else
    gnome-terminal -- bash -c "$0; exec bash"
fi

--

/home/xyz/workspace/LLAMA_CPP/9f2da5871f4bbd205b8a3b952cdc76283218d595/llama.cpp/build/bin/llama-server
Waiting for server to start...
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: Quadro M2000, compute capability 5.2, VMM: yes
build: 5276 (9f2da587) with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
system info: n_threads = 8, n_threads_batch = 8, total_threads = 32

system_info: n_threads = 8 (n_threads_batch = 8) / 32 | CUDA : ARCHS = 500,610,700,750,800 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | AARCH64_REPACK = 1 | 

main: binding port with default address family
main: HTTP server is listening, hostname: 0.0.0.0, port: 8001, http threads: 31
main: loading model
srv    load_model: loading model '/mnt/AI/LLM/DeepSeek-R1-UD-Q2_K_XL/DeepSeek-R1-UD-Q2_K_XL-00001-of-00005.gguf'
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 3090) - 17426 MiB free
llama_model_loader: additional 4 GGUFs metadata loaded.
llama_model_loader: loaded meta data with 48 key-value pairs and 1025 tensors from /mnt/AI/LLM/DeepSeek-R1-UD-Q2_K_XL/DeepSeek-R1-UD-Q2_K_XL-00001-of-00005.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              = deepseek2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = DeepSeek R1 BF16
llama_model_loader: - kv   3:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   4:                         general.size_label str              = 256x20B
llama_model_loader: - kv   5:                           general.repo_url str              = https://huggingface.co/unsloth
llama_model_loader: - kv   6:                      deepseek2.block_count u32              = 61
llama_model_loader: - kv   7:                   deepseek2.context_length u32              = 163840
llama_model_loader: - kv   8:                 deepseek2.embedding_length u32              = 7168
llama_model_loader: - kv   9:              deepseek2.feed_forward_length u32              = 18432
llama_model_loader: - kv  10:             deepseek2.attention.head_count u32              = 128
llama_model_loader: - kv  11:          deepseek2.attention.head_count_kv u32              = 128
llama_model_loader: - kv  12:                   deepseek2.rope.freq_base f32              = 10000.000000
llama_model_loader: - kv  13: deepseek2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  14:                deepseek2.expert_used_count u32              = 8
llama_model_loader: - kv  15:        deepseek2.leading_dense_block_count u32              = 3
llama_model_loader: - kv  16:                       deepseek2.vocab_size u32              = 129280
llama_model_loader: - kv  17:            deepseek2.attention.q_lora_rank u32              = 1536
llama_model_loader: - kv  18:           deepseek2.attention.kv_lora_rank u32              = 512
llama_model_loader: - kv  19:             deepseek2.attention.key_length u32              = 192
llama_model_loader: - kv  20:           deepseek2.attention.value_length u32              = 128
llama_model_loader: - kv  21:       deepseek2.expert_feed_forward_length u32              = 2048
llama_model_loader: - kv  22:                     deepseek2.expert_count u32              = 256
llama_model_loader: - kv  23:              deepseek2.expert_shared_count u32              = 1
llama_model_loader: - kv  24:             deepseek2.expert_weights_scale f32              = 2.500000
llama_model_loader: - kv  25:              deepseek2.expert_weights_norm bool             = true
llama_model_loader: - kv  26:               deepseek2.expert_gating_func u32              = 2
llama_model_loader: - kv  27:             deepseek2.rope.dimension_count u32              = 64
llama_model_loader: - kv  28:                deepseek2.rope.scaling.type str              = yarn
llama_model_loader: - kv  29:              deepseek2.rope.scaling.factor f32              = 40.000000
llama_model_loader: - kv  30: deepseek2.rope.scaling.original_context_length u32              = 4096
llama_model_loader: - kv  31: deepseek2.rope.scaling.yarn_log_multiplier f32              = 0.100000
llama_model_loader: - kv  32:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  33:                         tokenizer.ggml.pre str              = deepseek-v3
llama_model_loader: - kv  34:                      tokenizer.ggml.tokens arr[str,129280]  = ["<｜begin▁of▁sentence｜>", "<�...
llama_model_loader: - kv  35:                  tokenizer.ggml.token_type arr[i32,129280]  = [3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  36:                      tokenizer.ggml.merges arr[str,127741]  = ["Ġ t", "Ġ a", "i n", "Ġ Ġ", "h e...
llama_model_loader: - kv  37:                tokenizer.ggml.bos_token_id u32              = 0
llama_model_loader: - kv  38:                tokenizer.ggml.eos_token_id u32              = 1
llama_model_loader: - kv  39:            tokenizer.ggml.padding_token_id u32              = 128815
llama_model_loader: - kv  40:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  41:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  42:                    tokenizer.chat_template str              = {% if not add_generation_prompt is de...
llama_model_loader: - kv  43:               general.quantization_version u32              = 2
llama_model_loader: - kv  44:                          general.file_type u32              = 10
llama_model_loader: - kv  45:                                   split.no u16              = 0
llama_model_loader: - kv  46:                        split.tensors.count i32              = 1025
llama_model_loader: - kv  47:                                split.count u16              = 5
llama_model_loader: - type  f32:  361 tensors
llama_model_loader: - type q2_K:  171 tensors
llama_model_loader: - type q3_K:    3 tensors
llama_model_loader: - type q4_K:  306 tensors
llama_model_loader: - type q6_K:  184 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q2_K - Medium
print_info: file size   = 211.03 GiB (2.70 BPW) 
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
load: special tokens cache size = 819
load: token to piece cache size = 0.8223 MB
print_info: arch             = deepseek2
print_info: vocab_only       = 0
print_info: n_ctx_train      = 163840
print_info: n_embd           = 7168
print_info: n_layer          = 61
print_info: n_head           = 128
print_info: n_head_kv        = 128
print_info: n_rot            = 64
print_info: n_swa            = 0
print_info: n_swa_pattern    = 1
print_info: n_embd_head_k    = 192
print_info: n_embd_head_v    = 128
print_info: n_gqa            = 1
print_info: n_embd_k_gqa     = 24576
print_info: n_embd_v_gqa     = 16384
print_info: f_norm_eps       = 0.0e+00
print_info: f_norm_rms_eps   = 1.0e-06
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             = 18432
print_info: n_expert         = 256
print_info: n_expert_used    = 8
print_info: causal attn      = 1
print_info: pooling type     = 0
print_info: rope type        = 0
print_info: rope scaling     = yarn
print_info: freq_base_train  = 10000.0
print_info: freq_scale_train = 0.025
print_info: n_ctx_orig_yarn  = 4096
print_info: rope_finetuned   = unknown
print_info: ssm_d_conv       = 0
print_info: ssm_d_inner      = 0
print_info: ssm_d_state      = 0
print_info: ssm_dt_rank      = 0
print_info: ssm_dt_b_c_rms   = 0
print_info: model type       = 671B
print_info: model params     = 671.03 B
print_info: general.name     = DeepSeek R1 BF16
print_info: n_layer_dense_lead   = 3
print_info: n_lora_q             = 1536
print_info: n_lora_kv            = 512
print_info: n_embd_head_k_mla    = 0
print_info: n_embd_head_v_mla    = 0
print_info: n_ff_exp             = 2048
print_info: n_expert_shared      = 1
print_info: expert_weights_scale = 2.5
print_info: expert_weights_norm  = 1
print_info: expert_gating_func   = sigmoid
print_info: rope_yarn_log_mul    = 0.1000
print_info: vocab type       = BPE
print_info: n_vocab          = 129280
print_info: n_merges         = 127741
print_info: BOS token        = 0 '<｜begin▁of▁sentence｜>'
print_info: EOS token        = 1 '<｜end▁of▁sentence｜>'
print_info: EOT token        = 1 '<｜end▁of▁sentence｜>'
print_info: PAD token        = 128815 '<｜PAD▁TOKEN｜>'
print_info: LF token         = 201 'Ċ'
print_info: FIM PRE token    = 128801 '<｜fim▁begin｜>'
print_info: FIM SUF token    = 128800 '<｜fim▁hole｜>'
print_info: FIM MID token    = 128802 '<｜fim▁end｜>'
print_info: EOG token        = 1 '<｜end▁of▁sentence｜>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = false)
load_tensors: offloading 0 repeating layers to GPU
load_tensors: offloaded 0/62 layers to GPU
load_tensors:    CUDA_Host model buffer size = 215601.95 MiB
load_tensors:          CPU model buffer size =   497.11 MiB
....................................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 16384
llama_context: n_ctx_per_seq = 16384
llama_context: n_batch       = 4096
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 0
llama_context: freq_base     = 10000.0
llama_context: freq_scale    = 0.025
llama_context: n_ctx_per_seq (16384) < n_ctx_train (163840) -- the full capacity of the model will not be utilized
llama_context:        CPU  output buffer size =     0.49 MiB
llama_kv_cache_unified: kv_size = 16384, type_k = 'q4_0', type_v = 'f16', n_layer = 61, can_shift = 1, padding = 32
llama_kv_cache_unified:        CPU KV buffer size = 44408.00 MiB
llama_kv_cache_unified: KV self size  = 44408.00 MiB, K (q4_0): 13176.00 MiB, V (f16): 31232.00 MiB
llama_context:      CUDA0 compute buffer size =  5017.50 MiB
llama_context:  CUDA_Host compute buffer size =   112.01 MiB
llama_context: graph nodes  = 4842
llama_context: graph splits = 1148 (with bs=512), 1 (with bs=1)
common_init_from_params: setting dry_penalty_last_n to ctx_size = 16384
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 = 16384
main: model loaded
main: chat template, chat_template: {% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% set ns = namespace(is_first=false, is_tool=false, is_output_first=true, system_prompt='', is_first_sp=true) %}{%- for message in messages %}{%- if message['role'] == 'system' %}{%- if ns.is_first_sp %}{% set ns.system_prompt = ns.system_prompt + message['content'] %}{% set ns.is_first_sp = false %}{%- else %}{% set ns.system_prompt = ns.system_prompt + '\n\n' + message['content'] %}{%- endif %}{%- endif %}{%- endfor %}{{ bos_token }}{{ ns.system_prompt }}{%- for message in messages %}{%- if message['role'] == 'user' %}{%- set ns.is_tool = false -%}{{'<｜User｜>' + message['content']}}{%- endif %}{%- if message['role'] == 'assistant' and 'tool_calls' in message %}{%- set ns.is_tool = false -%}{%- for tool in message['tool_calls'] %}{%- if not ns.is_first %}{%- if message['content'] is none %}{{'<｜Assistant｜><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\n' + '' + '\n' + tool['function']['arguments'] + '\n' + '' + '<｜tool▁call▁end｜>'}}{%- else %}{{'<｜Assistant｜>' + message['content'] + '<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\n' + '' + '\n' + tool['function']['arguments'] + '\n' + '' + '<｜tool▁call▁end｜>'}}{%- endif %}{%- set ns.is_first = true -%}{%- else %}{{'\n' + '<｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\n' + '' + '\n' + tool['function']['arguments'] + '\n' + '' + '<｜tool▁call▁end｜>'}}{%- endif %}{%- endfor %}{{'<｜tool▁calls▁end｜><｜end▁of▁sentence｜>'}}{%- endif %}{%- if message['role'] == 'assistant' and 'tool_calls' not in message %}{%- if ns.is_tool %}{{'<｜tool▁outputs▁end｜>' + message['content'] + '<｜end▁of▁sentence｜>'}}{%- set ns.is_tool = false -%}{%- else %}{% set content = message['content'] %}{% if '</think>' in content %}{% set content = content.split('</think>')[-1] %}{% endif %}{{'<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>'}}{%- endif %}{%- endif %}{%- if message['role'] == 'tool' %}{%- set ns.is_tool = true -%}{%- if ns.is_output_first %}{{'<｜tool▁outputs▁begin｜><｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- set ns.is_output_first = false %}{%- else %}{{'<｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- endif %}{%- endif %}{%- endfor -%}{% if ns.is_tool %}{{'<｜tool▁outputs▁end｜>'}}{% endif %}{% if add_generation_prompt and not ns.is_tool %}{{'<｜Assistant｜>'}}{% endif %}, example_format: 'You are a helpful assistant

<｜User｜>Hello<｜Assistant｜>Hi there<｜end▁of▁sentence｜><｜User｜>How are you?<｜Assistant｜>'
main: server is listening on http://0.0.0.0:8001 - starting the main loop

[MB7979]

Might be related to the issues many of us who are doing partial offloading are having since the MLA commit.

For me the gibberish began with the MLA commit, but I was able to somewhat workaround it by enabling Q8_0 k cache. I see you have Q4_0 cache enabled so perhaps you are only just hitting it now for whatever reason with newer changes.

[JohannesGaessler]

Calculating perplexity on Wikitext using Deepseek V2 Lite q4_0 with max. GPU layers:

[1]3.9466,[2]5.4965,[3]5.1995,[4]5.6518,[5]6.1076,[6]6.8265,[7]7.0578,[8]7.4052,[9]7.8312,[10]8.1759,

With -ngl 0:

[1]nan,[2]nan,[3]nan,[4]nan,[5]nan,[6]nan,[7]nan,[8]nan,[9]nan,[10]nan,

With -ngl 23:

[1]nan,[2]nan,[3]nan,[4]nan,[5]nan,[6]nan,[7]nan,[8]nan,[9]nan,[10]nan,

With -ngl 24:

[1]3.9466,[2]5.4965,[3]5.1995,[4]5.6518,[5]6.1076,[6]6.8265,[7]nan,[8]nan,[9]nan,[10]nan,

With -ngl 25:

[1]3.9466,[2]5.4965,[3]5.1995,[4]5.6518,[5]6.1076,[6]6.8265,[7]7.0578,[8]7.4052,[9]7.8312,[10]8.1759,

Without a GPU at all:

[1]3.9699,[2]5.5112,[3]5.2249,[4]5.6893,[5]6.1456,[6]6.8579,[7]7.0851,[8]7.4328,[9]7.8667,[10]8.2129,

So it does look like there are issues with partial offloading.

[danielhanchen]

Oh yes I was supposed to bring the discussion over from https://huggingface.co/unsloth/DeepSeek-V3-0324-GGUF-UD/discussions/2 - thanks @joesixpaq for the extensive tests.

I think the MLA commit might still be having some issues - the new MMQ commit @JohannesGaessler made I think should be fixed with some small fixes, but unsure if there are some interactions going on.

The only thing I can confirm for now is full GPU offloading (ie all layers) seem to work OK - most people are having gibberish outputs when CPU offloading is done

[joesixpaq]

The problem disappeared when I specified --no-kv-offload in this first broken commit e1e8e09.

In another try, I additionally offloaded 5 layers to GPU (max for RTX 3090), and it worked as well.

[joesixpaq]

It works with the latest commit 66645a5 and --no-kv-offload.

/home/ai/workspace/LLAMA_CPP/66645a5285d8c4c5f9a3b3f360d042baac2d820a/llama.cpp/build/bin/llama-server
Waiting for server to start...
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: Quadro M2000, compute capability 5.2, VMM: yes
build: 5281 (66645a5) with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
system info: n_threads = 8, n_threads_batch = 8, total_threads = 32

system_info: n_threads = 8 (n_threads_batch = 8) / 32 | CUDA : ARCHS = 500,610,700,750,800 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | AARCH64_REPACK = 1 |

main: binding port with default address family
main: HTTP server is listening, hostname: 0.0.0.0, port: 8000, http threads: 31
main: loading model
srv load_model: loading model '/mnt/AI/LLM/DeepSeek-R1-UD-Q2_K_XL/DeepSeek-R1-UD-Q2_K_XL-00001-of-00005.gguf'
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 3090) - 23873 MiB free
llama_model_loader: additional 4 GGUFs metadata loaded.
llama_model_loader: loaded meta data with 48 key-value pairs and 1025 tensors from /mnt/AI/LLM/DeepSeek-R1-UD-Q2_K_XL/DeepSeek-R1-UD-Q2_K_XL-00001-of-00005.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 = deepseek2
llama_model_loader: - kv 1: general.type str = model
llama_model_loader: - kv 2: general.name str = DeepSeek R1 BF16
llama_model_loader: - kv 3: general.quantized_by str = Unsloth
llama_model_loader: - kv 4: general.size_label str = 256x20B
llama_model_loader: - kv 5: general.repo_url str = https://huggingface.co/unsloth
llama_model_loader: - kv 6: deepseek2.block_count u32 = 61
llama_model_loader: - kv 7: deepseek2.context_length u32 = 163840
llama_model_loader: - kv 8: deepseek2.embedding_length u32 = 7168
llama_model_loader: - kv 9: deepseek2.feed_forward_length u32 = 18432
llama_model_loader: - kv 10: deepseek2.attention.head_count u32 = 128
llama_model_loader: - kv 11: deepseek2.attention.head_count_kv u32 = 128
llama_model_loader: - kv 12: deepseek2.rope.freq_base f32 = 10000.000000
llama_model_loader: - kv 13: deepseek2.attention.layer_norm_rms_epsilon f32 = 0.000001
llama_model_loader: - kv 14: deepseek2.expert_used_count u32 = 8
llama_model_loader: - kv 15: deepseek2.leading_dense_block_count u32 = 3
llama_model_loader: - kv 16: deepseek2.vocab_size u32 = 129280
llama_model_loader: - kv 17: deepseek2.attention.q_lora_rank u32 = 1536
llama_model_loader: - kv 18: deepseek2.attention.kv_lora_rank u32 = 512
llama_model_loader: - kv 19: deepseek2.attention.key_length u32 = 192
llama_model_loader: - kv 20: deepseek2.attention.value_length u32 = 128
llama_model_loader: - kv 21: deepseek2.expert_feed_forward_length u32 = 2048
llama_model_loader: - kv 22: deepseek2.expert_count u32 = 256
llama_model_loader: - kv 23: deepseek2.expert_shared_count u32 = 1
llama_model_loader: - kv 24: deepseek2.expert_weights_scale f32 = 2.500000
llama_model_loader: - kv 25: deepseek2.expert_weights_norm bool = true
llama_model_loader: - kv 26: deepseek2.expert_gating_func u32 = 2
llama_model_loader: - kv 27: deepseek2.rope.dimension_count u32 = 64
llama_model_loader: - kv 28: deepseek2.rope.scaling.type str = yarn
llama_model_loader: - kv 29: deepseek2.rope.scaling.factor f32 = 40.000000
llama_model_loader: - kv 30: deepseek2.rope.scaling.original_context_length u32 = 4096
llama_model_loader: - kv 31: deepseek2.rope.scaling.yarn_log_multiplier f32 = 0.100000
llama_model_loader: - kv 32: tokenizer.ggml.model str = gpt2
llama_model_loader: - kv 33: tokenizer.ggml.pre str = deepseek-v3
llama_model_loader: - kv 34: tokenizer.ggml.tokens arr[str,129280] = ["<｜begin▁of▁sentence｜>", "<�...
llama_model_loader: - kv 35: tokenizer.ggml.token_type arr[i32,129280] = [3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 36: tokenizer.ggml.merges arr[str,127741] = ["Ġ t", "Ġ a", "i n", "Ġ Ġ", "h e...
llama_model_loader: - kv 37: tokenizer.ggml.bos_token_id u32 = 0
llama_model_loader: - kv 38: tokenizer.ggml.eos_token_id u32 = 1
llama_model_loader: - kv 39: tokenizer.ggml.padding_token_id u32 = 128815
llama_model_loader: - kv 40: tokenizer.ggml.add_bos_token bool = true
llama_model_loader: - kv 41: tokenizer.ggml.add_eos_token bool = false
llama_model_loader: - kv 42: tokenizer.chat_template str = {% if not add_generation_prompt is de...
llama_model_loader: - kv 43: general.quantization_version u32 = 2
llama_model_loader: - kv 44: general.file_type u32 = 10
llama_model_loader: - kv 45: split.no u16 = 0
llama_model_loader: - kv 46: split.tensors.count i32 = 1025
llama_model_loader: - kv 47: split.count u16 = 5
llama_model_loader: - type f32: 361 tensors
llama_model_loader: - type q2_K: 171 tensors
llama_model_loader: - type q3_K: 3 tensors
llama_model_loader: - type q4_K: 306 tensors
llama_model_loader: - type q6_K: 184 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type = Q2_K - Medium
print_info: file size = 211.03 GiB (2.70 BPW)
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
load: special tokens cache size = 819
load: token to piece cache size = 0.8223 MB
print_info: arch = deepseek2
print_info: vocab_only = 0
print_info: n_ctx_train = 163840
print_info: n_embd = 7168
print_info: n_layer = 61
print_info: n_head = 128
print_info: n_head_kv = 128
print_info: n_rot = 64
print_info: n_swa = 0
print_info: n_swa_pattern = 1
print_info: n_embd_head_k = 192
print_info: n_embd_head_v = 128
print_info: n_gqa = 1
print_info: n_embd_k_gqa = 24576
print_info: n_embd_v_gqa = 16384
print_info: f_norm_eps = 0.0e+00
print_info: f_norm_rms_eps = 1.0e-06
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 = 18432
print_info: n_expert = 256
print_info: n_expert_used = 8
print_info: causal attn = 1
print_info: pooling type = 0
print_info: rope type = 0
print_info: rope scaling = yarn
print_info: freq_base_train = 10000.0
print_info: freq_scale_train = 0.025
print_info: n_ctx_orig_yarn = 4096
print_info: rope_finetuned = unknown
print_info: ssm_d_conv = 0
print_info: ssm_d_inner = 0
print_info: ssm_d_state = 0
print_info: ssm_dt_rank = 0
print_info: ssm_dt_b_c_rms = 0
print_info: model type = 671B
print_info: model params = 671.03 B
print_info: general.name = DeepSeek R1 BF16
print_info: n_layer_dense_lead = 3
print_info: n_lora_q = 1536
print_info: n_lora_kv = 512
print_info: n_embd_head_k_mla = 0
print_info: n_embd_head_v_mla = 0
print_info: n_ff_exp = 2048
print_info: n_expert_shared = 1
print_info: expert_weights_scale = 2.5
print_info: expert_weights_norm = 1
print_info: expert_gating_func = sigmoid
print_info: rope_yarn_log_mul = 0.1000
print_info: vocab type = BPE
print_info: n_vocab = 129280
print_info: n_merges = 127741
print_info: BOS token = 0 '<｜begin▁of▁sentence｜>'
print_info: EOS token = 1 '<｜end▁of▁sentence｜>'
print_info: EOT token = 1 '<｜end▁of▁sentence｜>'
print_info: PAD token = 128815 '<｜PAD▁TOKEN｜>'
print_info: LF token = 201 'Ċ'
print_info: FIM PRE token = 128801 '<｜fim▁begin｜>'
print_info: FIM SUF token = 128800 '<｜fim▁hole｜>'
print_info: FIM MID token = 128802 '<｜fim▁end｜>'
print_info: EOG token = 1 '<｜end▁of▁sentence｜>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = false)
load_tensors: offloading 0 repeating layers to GPU
load_tensors: offloaded 0/62 layers to GPU
load_tensors: CUDA_Host model buffer size = 215601.95 MiB
load_tensors: CPU model buffer size = 497.11 MiB
....................................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max = 1
llama_context: n_ctx = 16384
llama_context: n_ctx_per_seq = 16384
llama_context: n_batch = 4096
llama_context: n_ubatch = 512
llama_context: causal_attn = 1
llama_context: flash_attn = 0
llama_context: freq_base = 10000.0
llama_context: freq_scale = 0.025
llama_context: n_ctx_per_seq (16384) < n_ctx_train (163840) -- the full capacity of the model will not be utilized
llama_context: CPU output buffer size = 0.49 MiB
llama_kv_cache_unified: kv_size = 16384, type_k = 'q4_0', type_v = 'f16', n_layer = 61, can_shift = 1, padding = 32
llama_kv_cache_unified: CPU KV buffer size = 44408.00 MiB
llama_kv_cache_unified: KV self size = 44408.00 MiB, K (q4_0): 13176.00 MiB, V (f16): 31232.00 MiB
llama_context: CUDA0 compute buffer size = 2783.50 MiB
llama_context: CUDA_Host compute buffer size = 4256.01 MiB
llama_context: graph nodes = 4842
llama_context: graph splits = 1087 (with bs=512), 1 (with bs=1)
common_init_from_params: setting dry_penalty_last_n to ctx_size = 16384
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 = 16384
main: model loaded
main: chat template, chat_template: {% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% set ns = namespace(is_first=false, is_tool=false, is_output_first=true, system_prompt='', is_first_sp=true) %}{%- for message in messages %}{%- if message['role'] == 'system' %}{%- if ns.is_first_sp %}{% set ns.system_prompt = ns.system_prompt + message['content'] %}{% set ns.is_first_sp = false %}{%- else %}{% set ns.system_prompt = ns.system_prompt + '\n\n' + message['content'] %}{%- endif %}{%- endif %}{%- endfor %}{{ bos_token }}{{ ns.system_prompt }}{%- for message in messages %}{%- if message['role'] == 'user' %}{%- set ns.is_tool = false -%}{{'<｜User｜>' + message['content']}}{%- endif %}{%- if message['role'] == 'assistant' and 'tool_calls' in message %}{%- set ns.is_tool = false -%}{%- for tool in message['tool_calls'] %}{%- if not ns.is_first %}{%- if message['content'] is none %}{{'<｜Assistant｜><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\n' + 'json' + '\n' + tool['function']['arguments'] + '\n' + '' + '<｜tool▁call▁end｜>'}}{%- else %}{{'<｜Assistant｜>' + message['content'] + '<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\n' + 'json' + '\n' + tool['function']['arguments'] + '\n' + '' + '<｜tool▁call▁end｜>'}}{%- endif %}{%- set ns.is_first = true -%}{%- else %}{{'\n' + '<｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\n' + 'json' + '\n' + tool['function']['arguments'] + '\n' + '' + '<｜tool▁call▁end｜>'}}{%- endif %}{%- endfor %}{{'<｜tool▁calls▁end｜><｜end▁of▁sentence｜>'}}{%- endif %}{%- if message['role'] == 'assistant' and 'tool_calls' not in message %}{%- if ns.is_tool %}{{'<｜tool▁outputs▁end｜>' + message['content'] + '<｜end▁of▁sentence｜>'}}{%- set ns.is_tool = false -%}{%- else %}{% set content = message['content'] %}{% if '' in content %}{% set content = content.split('')[-1] %}{% endif %}{{'<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>'}}{%- endif %}{%- endif %}{%- if message['role'] == 'tool' %}{%- set ns.is_tool = true -%}{%- if ns.is_output_first %}{{'<｜tool▁outputs▁begin｜><｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- set ns.is_output_first = false %}{%- else %}{{'<｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- endif %}{%- endif %}{%- endfor -%}{% if ns.is_tool %}{{'<｜tool▁outputs▁end｜>'}}{% endif %}{% if add_generation_prompt and not ns.is_tool %}{{'<｜Assistant｜>'}}{% endif %}, example_format: 'You are a helpful assistant

<｜User｜>Hello<｜Assistant｜>Hi there<｜end▁of▁sentence｜><｜User｜>How are you?<｜Assistant｜>'
main: server is listening on http://0.0.0.0:8000 - starting the main loop
Server ready! Launching browser...
Opening in existing browser session.

[JohannesGaessler]

@danielhanchen since you're already here, can I recruit you to help with investigating #13287 (comment) ?

[danielhanchen]

I could try, but I doubt I could be helpful! I've only recently started coding on the outer layers of llama.cpp, but not any internals

[slaren]

So it does look like there are issues with partial offloading.

It seems to work with a BF16 model. Maybe some padding that is not cleared correctly when copying the quantized tensors to VRAM?

[JohannesGaessler]

Yes, I already figured out that the issue specifically occurs for the combination of MMQ and ne00 % 256 != 0.

[JohannesGaessler]

I could try, but I doubt I could be helpful! I've only recently started coding on the outer layers of llama.cpp, but not any internals

What I meant is, since you are already running extensive tests with perplexity and KL divergence, could you check whether my changes have made the model predictions worse in a statistically significant way?

[JohannesGaessler]

Should be fixed by #13320 .

[MB7979]

This fixed the issues I’ve been having. Thank you very much.

[danielhanchen]

@JohannesGaessler I was just checking all your new changes - great work as usual!

imatrix.cpp sadly still gets errors with:

CUDA error: invalid configuration argument
  current device: 0, in function ggml_cuda_mul_mat_id at llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:2062
  cudaGetLastError()

ie at:

get_rows_cuda(src1->data, src1->type, ids_to_sorted, src1_sorted.ptr, type_src1_sorted,
        ne10, nb11, nb12, nb13,
        ne_get_rows, 1, 1, sizeof(int32_t), ne_get_rows*sizeof(int32_t), ne_get_rows*sizeof(int32_t),
        ne10*ts_src1_sorted, ne_get_rows*ne10*ts_src1_sorted, ne_get_rows*ne10*ts_src1_sorted, stream);
    CUDA_CHECK(cudaGetLastError());

Using --ubatch-size 8192 causes the error to occur on Qwen 3 30B MoE. --ubatch-size 8191 works fine.

My suspicion is because CUDA I think requires arguments to be < INT_MAX It's because Qwen has 128 experts, 2048 in dim, so 8192 * 2048 * 128 = 2147483648 > 2147483647 (INT_MAX).

8191 * 2048 * 128 = 2147221504, so less than INT_MAX.

Ie one of the arguments:

ne10*ts_src1_sorted, ne_get_rows*ne10*ts_src1_sorted, ne_get_rows*ne10*ts_src1_sorted

is exceeding INT_MAX, thus causing CUDA to error out.

I'll make a new issue!