'2026/09 글 목록

'2026/09'에 해당되는 글 88건

11:26:17 kicad 공식 사용법
11:04:47 kicad 발음
2026.09.27 스타크래프트 ai의 시대
2026.09.26 llm 초기값
2026.09.26 피곤
2026.09.25 흐음..
2026.09.24 diffusion gemma 실행해보기
2026.09.23 diffusion gemma
2026.09.23 직접만든 LLM .. part 2
2026.09.23 max7219 8-Digit LED Display Drivers

kicad 발음 (0)	2026.09.28
kicad 사용법 (0)	2026.08.14
kicad mcp, flux.ai (0)	2026.01.09
Kicad - open source PCB / circuit program (11)	2010.07.08

kicad 공식 사용법 (0)	2026.09.28
kicad 사용법 (0)	2026.08.14
kicad mcp, flux.ai (0)	2026.01.09
Kicad - open source PCB / circuit program (11)	2010.07.08

스타크래프트 2 - DLC 노바의 비밀 임무 끝! (0)	2019.01.26
스타2 공허의 유산 에필로그 끝! (0)	2019.01.01
스타 2 공허의 유산 끝! (0)	2018.12.31
스타2 공허의 유산 시작! (0)	2018.12.30
스타2 군단의 심장(저그) 보통 난이도 완료 (0)	2018.12.29

주말도 피곤 (0)	2026.09.06
아 몰라 또 질러 ㅋㅋㅋ (0)	2026.08.31
피아노 줍줍 (0)	2026.08.30
개피곤 (0)	2026.08.29
와.. 설마 에어컨 고장? (2)	2026.08.24

ai를 이용한 개발 관련 좋은글 (0)	2026.06.29
새로운 마약 - AI (0)	2026.03.17
agi를 느꼈다? (0)	2026.02.24
바이브 코딩 ai 코딩의 광풍 (0)	2025.10.13
ai killed programmer (0)	2025.05.23

diffusion gemma 실행해보기

음.. 가장 최신은 아니지만

b11158 를 받아서 빌드된걸 보니 llama-diffusion-cli가 없다.

/mnt/Downloads/llama-b11158$ ./llama-
llama-batched-bench  llama-gemma3-cli     llama-mtmd-cli       llama-results
llama-bench          llama-gguf-split     llama-mtmd-debug     llama-server
llama-cli            llama-imatrix        llama-perplexity     llama-tokenize
llama-completion     llama-llava-cli      llama-quantize       llama-tts
llama-fit-params     llama-minicpmv-cli   llama-qwen2vl-cli

그래서 빌드해둔거 보니까 있네? 머지?

:/mnt/Downloads/llama.cpp/build/bin$ ./llama
llama                          llama-lookup
llama-batched                  llama-lookup-create
llama-batched-bench            llama-lookup-merge
llama-bench                    llama-lookup-stats
llama-cli                      llama-minicpmv-cli
llama-completion               llama-mtmd-cli
llama-convert-llama2c-to-ggml  llama-mtmd-debug
llama-cvector-generator        llama-parallel
llama-debug                    llama-passkey
llama-diffusion-cli            llama-perplexity
llama-embedding                llama-q8dot
llama-eval-callback            llama-quantize
llama-export-lora              llama-qwen2vl-cli
llama-finetune                 llama-results
llama-fit-params               llama-retrieval
llama-gemma3-cli               llama-server
llama-gen-docs                 llama-simple
llama-gguf                     llama-simple-chat
llama-gguf-hash                llama-speculative
llama-gguf-split               llama-speculative-simple
llama-idle                     llama-tokenize
llama-imatrix                  llama-tts
llama-llava-cli                llama-vdot
llama-lookahead

용량 자체는 동일 양자화면 차이가 거의 나지 않는다.

diffusiongemma-26B-A4B-it-Q4_K_M.gguf 16.8 GB

[링크 : https://huggingface.co/unsloth/diffusiongemma-26B-A4B-it-GGUF]

gemma-4-26B-A4B-it-UD-Q4_K_M.gguf 16.9 GB

[링크 : https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF]

근데 실행이 안된다?

/mnt/Downloads/model/dgemma4-26b$ /mnt/Downloads/llama.cpp/build/bin/llama-diffusion-cli -m ./diffusiongemma-26B-A4B-it-Q4_K_M.gguf

0.00.283.015 E llama_model_load: error loading model: unknown model architecture: 'diffusion-gemma'
0.00.283.023 E llama_model_load_from_file_impl: failed to load model
0.00.283.024 E error: failed to load model './diffusiongemma-26B-A4B-it-Q4_K_M.gguf'

으음? 봐도 모르겠다!?

/mnt/Downloads/model/dgemma4-26b$ /mnt/Downloads/llama.cpp/build/bin/llama-diffusion-cli -m ./diffusiongemma-26B-A4B-it-Q4_K_M.gguf -v
0.00.112.071 I llama_model_loader: loaded meta data with 44 key-value pairs and 692 tensors from ./diffusiongemma-26B-A4B-it-Q4_K_M.gguf (version GGUF V3 (latest))
0.00.112.117 I llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
0.00.112.125 I llama_model_loader: - kv   0:                       general.architecture str              = diffusion-gemma
0.00.112.126 I llama_model_loader: - kv   1:                               general.type str              = model
0.00.112.126 I llama_model_loader: - kv   2:                               general.name str              = Dg_Rc0P1_Patched
0.00.112.127 I llama_model_loader: - kv   3:                         general.size_label str              = 128x2.6B
0.00.112.130 I llama_model_loader: - kv   4:                diffusion-gemma.block_count u32              = 30
0.00.112.131 I llama_model_loader: - kv   5:             diffusion-gemma.context_length u32              = 262144
0.00.112.131 I llama_model_loader: - kv   6:           diffusion-gemma.embedding_length u32              = 2816
0.00.112.132 I llama_model_loader: - kv   7:        diffusion-gemma.feed_forward_length u32              = 2112
0.00.112.132 I llama_model_loader: - kv   8:       diffusion-gemma.attention.head_count u32              = 16
0.00.112.147 I llama_model_loader: - kv   9:    diffusion-gemma.attention.head_count_kv arr[i32,30]      = [8, 8, 8, 8, 8, 2, 8, 8, 8, 8, 8, 2, ...
0.00.112.151 I llama_model_loader: - kv  10:             diffusion-gemma.rope.freq_base f32              = 1000000.000000
0.00.112.152 I llama_model_loader: - kv  11:         diffusion-gemma.rope.freq_base_swa f32              = 10000.000000
0.00.112.153 I llama_model_loader: - kv  12: diffusion-gemma.attention.layer_norm_rms_epsilon f32              = 0.000001
0.00.112.153 I llama_model_loader: - kv  13:               diffusion-gemma.expert_count u32              = 128
0.00.112.154 I llama_model_loader: - kv  14:          diffusion-gemma.expert_used_count u32              = 8
0.00.112.155 I llama_model_loader: - kv  15:       diffusion-gemma.attention.key_length u32              = 512
0.00.112.155 I llama_model_loader: - kv  16:     diffusion-gemma.attention.value_length u32              = 512
0.00.112.156 I llama_model_loader: - kv  17:    diffusion-gemma.final_logit_softcapping f32              = 30.000000
0.00.112.157 I llama_model_loader: - kv  18:   diffusion-gemma.attention.sliding_window u32              = 1024
0.00.112.157 I llama_model_loader: - kv  19: diffusion-gemma.attention.shared_kv_layers u32              = 0
0.00.112.158 I llama_model_loader: - kv  20: diffusion-gemma.embedding_length_per_layer_input u32              = 0
0.00.112.161 I llama_model_loader: - kv  21: diffusion-gemma.attention.sliding_window_pattern arr[bool,30]     = [true, true, true, true, true, false,...
0.00.112.162 I llama_model_loader: - kv  22:   diffusion-gemma.attention.key_length_swa u32              = 256
0.00.112.162 I llama_model_loader: - kv  23: diffusion-gemma.attention.value_length_swa u32              = 256
0.00.112.163 I llama_model_loader: - kv  24: diffusion-gemma.expert_feed_forward_length u32              = 704
0.00.112.163 I llama_model_loader: - kv  25:       diffusion-gemma.rope.dimension_count u32              = 512
0.00.112.164 I llama_model_loader: - kv  26:   diffusion-gemma.rope.dimension_count_swa u32              = 256
0.00.112.164 I llama_model_loader: - kv  27:           diffusion-gemma.attention.causal bool             = false
0.00.112.165 I llama_model_loader: - kv  28:                    diffusion.canvas_length u32              = 256
0.00.112.165 I llama_model_loader: - kv  29:                       tokenizer.ggml.model str              = gemma4
0.00.144.043 I llama_model_loader: - kv  30:                      tokenizer.ggml.tokens arr[str,262144]  = ["<pad>", "<eos>", "<bos>", "<unk>", ...
0.00.216.722 I llama_model_loader: - kv  31:                      tokenizer.ggml.scores arr[f32,262144]  = [-1000.000000, -1000.000000, -1000.00...
0.00.223.967 I llama_model_loader: - kv  32:                  tokenizer.ggml.token_type arr[i32,262144]  = [3, 3, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, ...
0.00.278.765 I llama_model_loader: - kv  33:                      tokenizer.ggml.merges arr[str,514906]  = ["\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n \n", ...
0.00.278.769 I llama_model_loader: - kv  34:                tokenizer.ggml.bos_token_id u32              = 2
0.00.278.769 I llama_model_loader: - kv  35:                tokenizer.ggml.eos_token_id u32              = 1
0.00.278.770 I llama_model_loader: - kv  36:            tokenizer.ggml.unknown_token_id u32              = 3
0.00.278.770 I llama_model_loader: - kv  37:            tokenizer.ggml.padding_token_id u32              = 0
0.00.278.786 I llama_model_loader: - kv  38:               tokenizer.ggml.mask_token_id u32              = 4
0.00.278.787 I llama_model_loader: - kv  39:            tokenizer.ggml.add_space_prefix bool             = false
0.00.278.787 I llama_model_loader: - kv  40:               tokenizer.ggml.add_bos_token bool             = true
0.00.278.788 I llama_model_loader: - kv  41:               general.quantization_version u32              = 2
0.00.278.789 I llama_model_loader: - kv  42:                          general.file_type u32              = 15
0.00.278.792 I llama_model_loader: - kv  43:                    tokenizer.chat_template str              = {%- macro format_parameters(propertie...
0.00.278.795 I llama_model_loader: - type  f32:  423 tensors
0.00.278.797 I llama_model_loader: - type q5_0:   33 tensors
0.00.278.798 I llama_model_loader: - type q8_0:   28 tensors
0.00.278.798 I llama_model_loader: - type q4_K:  194 tensors
0.00.278.798 I llama_model_loader: - type q6_K:   14 tensors
0.00.278.801 I print_info: file format = GGUF V3 (latest)
0.00.278.802 I print_info: file type   = Q4_K - Medium
0.00.278.805 I print_info: file size   = 15.64 GiB (5.32 BPW)
0.00.282.795 E llama_model_load: error loading model: unknown model architecture: 'diffusion-gemma'
0.00.282.802 E llama_model_load_from_file_impl: failed to load model
0.00.282.802 E error: failed to load model './diffusiongemma-26B-A4B-it-Q4_K_M.gguf'

모델 카드랑 claude에서 해보니

정식으로 gemma-diffusion은 llama.cpp에서 채택이 안되고 특정 커밋으로 가야 한다고 한다.

그래도 돌릴 수 있는게 어디냐~

git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp

# check out the DiffusionGemma PR (#24423)
gh pr checkout 24423
# or, without the gh CLI:
# git fetch origin pull/24423/head:diffusiongemma && git checkout diffusiongemma

# build with CUDA (drop -DGGML_CUDA=ON for a CPU-only build)
cmake -B build -DGGML_CUDA=ON
cmake --build build -j --config Release --target llama-diffusion-cli

[링크 : https://huggingface.co/unsloth/diffusiongemma-26B-A4B-it-GGUF]

다시 보니 여기서도 diffusion model이라고 하지 gemma-diffusion이라고 하진 않았네

[링크 : https://github.com/ggml-org/llama.cpp/blob/master/examples/diffusion/README.md]

응 니꺼에서는 못돌려~

/mnt/Downloads/llama.cpp-dgemma/build/bin$ ./llama-diffusion-cli -m /mnt/Downloads/model/dgemma4-26b/diffusiongemma-26B-A4B-it-Q4_K_M.gguf -sm layer -ts 9,10\
>
0.00.773.112 W load: control-looking token:    212 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
0.00.773.902 W load: control-looking token:     50 '<|tool_response>' was not control-type; this is probably a bug in the model. its type will be overridden
0.00.816.334 W load: special_eog_ids contains '<|tool_response>', removing '</s>' token from EOG list
0.05.615.403 I diffusion: --fit has no effect here; context is sized from -n and the canvas. Set -ngl / --n-cpu-moe to control device memory.
0.05.620.890 W resolve_fused_ops: layer 5 is assigned to device CUDA0 but Flash Attention is assigned to device CPU (usually due to missing support)
0.05.620.893 W resolve_fused_ops: Flash Attention not supported, set to disabled
0.10.954.103 E ggml_backend_cuda_buffer_type_alloc_buffer: allocating 1343.01 MiB on device 0: cudaMalloc failed: out of memory
0.10.954.108 E ggml_gallocr_reserve_n_impl: failed to allocate CUDA0 buffer of size 1408245760
0.10.954.109 E graph_reserve: failed to allocate compute buffers
0.10.954.115 W sched_reserve: compute buffer allocation failed, retrying without pipeline parallelism
0.11.063.993 I diffusion_params: steps=128 schedule=0 algorithm=4 temperature=0.800 eps=0.001000 mask_token=4
0.11.064.183 I diffusion_eb: kv cache auto-off (2 GPUs; pass --diffusion-kv-cache on to force)
0.11.064.184 I diffusion_eb: gpu sampling off (2 GPUs; sc_dev is single-device)
0.11.064.184 I diffusion_eb: gpu sample reduce off (needs --diffusion-gpu-sampling on / sc_dev)
0.11.064.186 I diffusion_eb: max_steps=48 t=[0.400,0.800] entropy_bound=0.1000 stability=1 confidence=0.0050 kv_cache=off gpu_sampling=off sample_reduce=off
/mnt/Downloads/llama.cpp-dgemma/ggml/src/ggml-cuda/ggml-cuda.cu:109: CUDA error
0.11.340.213 E CUDA error: out of memory
0.11.340.217 E   current device: 0, in function alloc at /mnt/Downloads/llama.cpp-dgemma/ggml/src/ggml-cuda/ggml-cuda.cu:589
0.11.340.218 E   cuMemCreate(&handle, reserve_size, &prop, 0)
[New LWP 3338013]
[New LWP 3338014]
[New LWP 3338017]
[New LWP 3338018]
[New LWP 3338019]
[New LWP 3338020]
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
0x00007e23920ea3ef in __GI___wait4 (pid=3338039, stat_loc=0x0, options=0, usage=0x0) at ../sysdeps/unix/sysv/linux/wait4.c:30
30      ../sysdeps/unix/sysv/linux/wait4.c: No such file or directory.
#0  0x00007e23920ea3ef in __GI___wait4 (pid=3338039, stat_loc=0x0, options=0, usage=0x0) at ../sysdeps/unix/sysv/linux/wait4.c:30
30      in ../sysdeps/unix/sysv/linux/wait4.c
#1  0x00007e2393183e3b in ggml_print_backtrace () from libggml-base.so.0
#2  0x00007e2393183fd2 in ggml_abort () from libggml-base.so.0
#3  0x00007e238b87ad46 in ggml_cuda_error(char const*, char const*, char const*, int, char const*) () from libggml-cuda.so.0
#4  0x00007e238b8928d3 in ggml_cuda_pool_vmm::alloc(unsigned long, unsigned long*) () from libggml-cuda.so.0
#5  0x00007e238b889fab in ggml_cuda_mul_mat_cublas(ggml_backend_cuda_context&, ggml_tensor const*, ggml_tensor const*, ggml_tensor*) () from libggml-cuda.so.0
#6  0x00007e238b88f849 in ggml_backend_cuda_graph_compute(ggml_backend*, ggml_cgraph*) () from libggml-cuda.so.0
#7  0x00007e23931a3d57 in ggml_backend_sched_graph_compute_async () from libggml-base.so.0
#8  0x00007e23929ec7b1 in llama_context::graph_compute(ggml_cgraph*, bool) () from libllama.so.0
#9  0x00007e23929eef65 in llama_context::process_ubatch(llama_ubatch const&, llm_graph_type, llama_memory_context_i*, ggml_status&) () from libllama.so.0
#10 0x00007e23929f6c6a in llama_context::encode(llama_batch const&) () from libllama.so.0
#11 0x00007e23929f90a0 in llama_decode () from libllama.so.0
#12 0x000057ed5b33e814 in diffusion_generate_entropy_bound(llama_context*, int const*, int*, int, diffusion_eb_params const&, int&) ()
#13 0x000057ed5b32f837 in main::{lambda(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&)#7}::operator()(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&) const ()
#14 0x000057ed5b32db9c in main ()
[Inferior 1 (process 3338011) detached]
Aborted (core dumped)

클로드 붙여서 어떻게든 돌려보려고 발악(!)

/mnt/Downloads/llama.cpp-dgemma/build/bin$ ./llama-diffusion-cli -m /mnt/Downloads/model/dgemma4-26b/diffusiongemma-26B-A4B-it-Q4_K_M.gguf   -ngl all -sm layer -ts 1,1 -ncmoe 6 -fit off -c 4096 -n 512 -p "Explain what a GPU is in two sentences."
0.00.784.683 W load: control-looking token:    212 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
0.00.785.427 W load: control-looking token:     50 '<|tool_response>' was not control-type; this is probably a bug in the model. its type will be overridden
0.00.828.578 W load: special_eog_ids contains '<|tool_response>', removing '</s>' token from EOG list
0.00.858.267 W llama_model_loader: tensor overrides to CPU are used with mmap enabled - consider using --load-mode none for better performance
0.05.288.187 I diffusion: -n 512 -> 2 blocks, n_ubatch=2560 n_batch=2560 n_ctx=4096 (canvas_length=256)
0.05.293.211 W resolve_fused_ops: layer 5 is assigned to device CUDA0 but Flash Attention is assigned to device CPU (usually due to missing support)
0.05.293.215 W resolve_fused_ops: Flash Attention not supported, set to disabled
0.10.920.964 I diffusion_params: steps=128 schedule=0 algorithm=4 temperature=0.800 eps=0.001000 mask_token=4
0.10.921.170 I diffusion_eb: kv cache auto-off (2 GPUs; pass --diffusion-kv-cache on to force)
0.10.921.172 I diffusion_eb: gpu sampling off (2 GPUs; sc_dev is single-device)
0.10.921.172 I diffusion_eb: gpu sample reduce off (needs --diffusion-gpu-sampling on / sc_dev)
0.10.921.174 I diffusion_eb: max_steps=48 t=[0.400,0.800] entropy_bound=0.1000 stability=1 confidence=0.0050 kv_cache=off gpu_sampling=off sample_reduce=off
diffusion step: 2/48 [==                                                ] 4%0.32.981.464 I I
diffusion step: 4/48 [====                                              ] 8%0.35.566.854 I
<|channel>thought
*   Topic: GPU (Graphics Processing Unit).
    *   Constraint: Exactly two sentences.
    *   Goal: Explain what it is.

    *   What is it? A specialized electronic circuit.
    *   What does it do? Handles image rendering, video, and parallel processing tasks.
    *   How does it differ from a CPU? Many cores optimized for speed.

    A GPU, or graphics processing unit, is a specialized processor designed to rapidly accelerate the creation and rendering of images for a display. Unlike a standard CPU, it uses many parallel cores to handle multiple complex tasks simultaneously, making it ideal for video editing and artificial intelligence.

    *   Sentence 1: A GPU is a specialized hardware processor designed to rapidly accelerate the creation and rendering of images and video.
    *   Sentence 2: Unlike a standard CPU, it uses many parallel cores to handle multiple complex tasks simultaneously, making it highly effective for gaming and artificial intelligence workloads.

    *   Two sentences? Yes.
    *   Accurate? Yes.<channel|>A GPU, or graphics processing unit, is a specialized hardware processor designed to rapidly accelerate the creation and rendering of images and video. Unlike a standard CPU, it uses many parallel cores to handle multiple complex tasks simultaneously, making it highly effective for gaming and artificial intelligence.
total time: 24659.24ms, time per step: 1072.14ms (23 steps over 2 blocks, entropy-bound)
throughput: 20.8 tok/s (512 tok in 24659.24ms), in-step parallel 239 tok/s (256-tok canvas x 11.5 steps/block)

context도 확 줄여도 메모리 엄청 먹나보다. 둘 다 거의 100% 때려버리네

뭔가.. 기본 질문이라도 있나? 프롬프트 없이 돌려도 이러고 난리다

/mnt/Downloads/llama.cpp-dgemma/build/bin$ ./llama-diffusion-cli -m /mnt/Downloads/model/dgemma4-26b/diffusiongemma-26B-A4B-it-Q4_K_M.gguf   -ngl all -sm layer -ts 1,1 -ncmoe 6 -fit off -c 4096 -n 512
0.00.771.302 W load: control-looking token:    212 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
0.00.772.085 W load: control-looking token:     50 '<|tool_response>' was not control-type; this is probably a bug in the model. its type will be overridden
0.00.815.021 W load: special_eog_ids contains '<|tool_response>', removing '</s>' token from EOG list
0.00.844.927 W llama_model_loader: tensor overrides to CPU are used with mmap enabled - consider using --load-mode none for better performance
0.05.277.471 I diffusion: -n 512 -> 2 blocks, n_ubatch=2560 n_batch=2560 n_ctx=4096 (canvas_length=256)
0.05.282.401 W resolve_fused_ops: layer 5 is assigned to device CUDA0 but Flash Attention is assigned to device CPU (usually due to missing support)
0.05.282.404 W resolve_fused_ops: Flash Attention not supported, set to disabled
0.10.802.856 I diffusion_params: steps=128 schedule=0 algorithm=4 temperature=0.800 eps=0.001000 mask_token=4
0.10.803.053 I diffusion_eb: kv cache auto-off (2 GPUs; pass --diffusion-kv-cache on to force)
0.10.803.055 I diffusion_eb: gpu sampling off (2 GPUs; sc_dev is single-device)
0.10.803.055 I diffusion_eb: gpu sample reduce off (needs --diffusion-gpu-sampling on / sc_dev)
0.10.803.057 I diffusion_eb: max_steps=48 t=[0.400,0.800] entropy_bound=0.1000 stability=1 confidence=0.0050 kv_cache=off gpu_sampling=off sample_reduce=off
diffusion step: 9/48 [=========                                         ] 18%0.45.079.364 I
<|channel>thought
The user has provided an image and wants me to describe or discuss it.

1.  **Analyze the image:**
    *   It's a close-up of a person's face.
    *   The person is smiling warmly.
    *   They are looking slightly upwards and to the side of the camera.
    *   The background is blurred (bokeh effect).
    *   The lighting is soft and warm, highlighting the contours of the face.
    *   The person looks like they could be of middle age or slightly older.
    *   They have dark hair and brown eyes.
    *   The overall mood is positive, hopeful, or joyful.

2.  **Formulate a description:**
    *   Identify the main subject: A person.
    *   Describe the action/expression: Smiling warmly, looking upwards.
    *   Describe the composition: Close-up shot, shallow depth of field.
    *   Describe the lighting: Warm, soft.
    *   Describe the mood: Happy, peaceful, serene.

3.  **Draft the final response:**
    *   Start with a general statement about what is in the image.
    *   Add details about the person's expression.
    *   Mention the photographic style (lighting, focus).<channel|>This is a close-up photograph of a person smiling warmly. They are looking slightly upwards and to the side with an expression that suggests hope or joy. The lighting is soft and warm, and the background is blurred, which keeps the focus entirely on the person's expression.
total time: 34290.01ms, time per step: 1071.56ms (32 steps over 2 blocks, entropy-bound)
throughput: 14.9 tok/s (512 tok in 34290.01ms), in-step parallel 239 tok/s (256-tok canvas x 16.0 steps/block)

아래 옵션으로 하니 먼가 되기 한다. ai님 만세!

./llama-diffusion-cli \
    -m /mnt/Downloads/model/dgemma4-26b/diffusiongemma-26B-A4B-it-Q4_K_M.gguf \
    -ngl all -sm layer -ts 58,42 -ncmoe 10 -fit off -c 4096 -n 1024 \
    -cnv --diffusion-visual --diffusion-visual-progress

일단 아까 설정한 -n 1024 로 한번에 1024 토큰 생성하는 거라서 그런가

답변이 넘어가면 diffusion step 이 다시 0/... 으로 돌아가버린다.

생성 토큰이 짧게 제한되서 그런가 답변이 나오다 끊어진 느낌

그래도 좀 신선하네 ㅎㅎ

> 파이썬으로 셀레니움을 통해 웹을 서칭하고 텍스트만 추출하고 makrdown 으로 변환후  md 파일과 pdf로 저장하는 기능을 구현해줘
<|channel>thought
Implement a Python script using Selenium to search the web, extract text, convert it to Markdown, and save it as both `.md` and `.pdf` files.
Python.
Selenium (for web automation/scraping), BeautifulSoup (usually helpful for parsing HTML), a Markdown conversion library (like `markdownify`), and a PDF generation library (like `fpdf2` or `pdf2kit`).

    *   *Web Scraping:* Selenium is requested. Needs a driver (Chrome is standard).
    *   *Text Extraction:* BeautifulSoup is better for cleaning HTML than just raw Selenium selectors.
    *   *Markdown Conversion:* `markdownify` is a popular choice for this.
    *   *PDF Generation:* `fpdf2` is straightforward for text-based PDFs. `pdf2kit` is better if converting from HTML/Markdown directly. Let's go with `fpdf2` for simplicity or mention the requirements.

    *   *Step 1: Install Libraries.* Need `selenium`, `beautifulsoup4`, `markdownify`, `fpdf2`.
    *   *Step 2: Setup Selenium.* Initialize Chrome driver.
    *   *Step 3: Search Logic.* Go to a search engine (e.g., Google), type a query, hit enter.
    *   *Step 4: Extraction.* Get page content (titles, paragraphs).
    *   *Step 5: Markdown Conversion.* Use `markdownify`.
    *   *Step 6: PDF Conversion.* Use `fpdf2`.

    *   *Imports:*
        ```python
        from selenium import webdriver
        from seleniuhttp://m.webdriver.chrome.service import Service
        from seleniuhttp://m.webdriver.common.by import By
        from seleniuhttp://m.webdriver.common.keys import Keys
        from bs4 import BeautifulSoup
        from markdownify import markdownify
        from fpdf import FPDF
        import time
        ```

    *   *Selenium Setup:*
        ```python
        options = webdriver.ChromeOptions()
        # options.add_argument('--headless') # Run without window
        driver = webdriver.Chrome(options=options)
        ```

    *   *Scraping Logic:*
        ```python
        driver.get(f"https://www.google.com/search?q={query}")
        # Wait for load
        html = driver.page_source
        soup = BeautifulSoup(html, 'html.parser')
        # Extract specific tags (h3, p)
        ```

    *   *Markdown Logic:*
        ```python
        md_text = markdownify(str(soup.body))
        ```

    *   *PDF Logic:*
        ```python
        pdf = FPDF()
        pdf.add_page()
        pdf.set_font("Arial", size=12) # Note: Korean needs a custom font!
        # Need to handle Korean characters in PDF.
        ```

    *   *Handling Korean Text:* Standard PDF fonts in FPDF don't support Korean by default. I must mention that the user needs a `.ttf` font (like NanumGothic).

    *   *Structure:*
        1.  Required libraries.
        2.  Implementation code.
        3.  Explanation of each part.
        4.  Important notes (Font issues, Driver setup).

    *   *Library check:* `pip install selenium beautifulsoup4 markdownify fpdf2`.
    *   *Code refinement:*
        *   Use a class or clear functions for structure.
        *   Add error handling (basic).
        *   Ensure the PDF part handles Unicode encoding for Korean.<channel|>파이썬의 **Selenium**을 사용하여 웹을 검색하고, 추출된 텍스트를 **Markdown**으로 변환한 뒤, 다시 **PDF** 파일로 저장하는 전체 과정을 구현해 드립니다.

### 1. 필요한 라이브러리 설치

터미널(또는 CMD)에서 아래 명령어를 입력하여 필요한 라이브러리를 설치해 주세요.

```bash
pip install selenium beautifulsoup4 markdownify fpdf2
```

### 2. 파이썬 코드 구현

이 코드는 구글 검색을 예시로 하며, PDF 생성 시 한글 깨짐 방지를 위해 **한글 폰트(.ttf)**가 필요합니다.

```python
import time
from selenium import webdriver
from seleniuhttp://m.webdriver.chrome.service import Service
from seleniuhttp://m.webdriver.chrome.options import Options
from seleniuhttp://m.webdriver.common.by import By
from seleniuhttp://m.webdriver.common.keys import Keys
from bs4 import BeautifulSoup
from markdownify import markdownify
from fpdf import FPDF
import os

class WebScraperToFiles:
    def
total time: 165742.13ms, time per step: 1883.43ms (88 steps over 4 blocks, entropy-bound)
throughput: 6.2 tok/s (1024 tok in 165742.13ms), in-step parallel 136 tok/s (256-tok canvas x 22.0 steps/block)

저작자표시 (새창열림)

'모종의 음모 > ai 프로그램' 카테고리의 다른 글

llm 초기값 (0)	2026.09.26
diffusion gemma (0)	2026.09.23
직접만든 LLM .. part 2 (0)	2026.09.23
wikimatrix (0)	2026.09.22
직접 LLM 만들...기? (0)	2026.09.22

Posted by 구차니

diffusion gemma

unsloth 형님들 만세!

3달전에 올라온 따끈한 모델

[링크 : https://github.com/ggml-org/llama.cpp/blob/master/examples/diffusion/README.md]

[링크 : https://huggingface.co/google/diffusiongemma-26B-A4B-it]

[링크 : https://huggingface.co/unsloth/diffusiongemma-26B-A4B-it-GGUF]

저작자표시 (새창열림)

'모종의 음모 > ai 프로그램' 카테고리의 다른 글

llm 초기값 (0)	2026.09.26
diffusion gemma 실행해보기 (0)	2026.09.24
직접만든 LLM .. part 2 (0)	2026.09.23
wikimatrix (0)	2026.09.22
직접 LLM 만들...기? (0)	2026.09.22

Posted by 구차니

직접만든 LLM .. part 2

어우 도대체 프론티어 모델들을 돌리는 데이터는 얼마나 어마어마하게 쌓여있어야 하고

얼마나 대단한 컴퓨팅 파워를 필요로 하는걸까?

대답은 이 난리가 나고 있다.

/mnt/llm-ko/3_small/gguf$ /mnt/Downloads/llama-b10488/llama-cli -m base42m-it-q4_0.gguf

Loading model...

▄▄ ▄▄
██ ██
██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
                                    ██    ██
                                    ▀▀    ▀▀

build      : b10488-9d77fa172
model      : base42m-it-q4_0.gguf
ftype      : Q4_0
modalities : text

available commands:
  /exit or Ctrl+C     stop or exit
  /regen              regenerate the last response
  /clear              clear the chat history
  /read <file>        add a text file
  /glob <pattern>     add text files using globbing pattern

> 안녕?
죄송해요, 저는 실시간 검색은 할 수 없어요. 제가 알기 어려워요

[ Prompt: 745.2 t/s | Generation: 868.1 t/s ]

> 서울은 어디야?
서울은 부산입니다

[ Prompt: 2201.1 t/s | Generation: 719.6 t/s ]

> 대한민국 수도는?
서울이에요

[ Prompt: 888.6 t/s | Generation: 543.7 t/s ]

> 서울이 먼데?
서울은 서울이에요

[ Prompt: 648.3 t/s | Generation: 268.8 t/s ]

> 넌 누구야?
서울이에요

[ Prompt: 420.3 t/s | Generation: 172.6 t/s ]

> 마이크로소프트는?
Microsoft Windows에서 작동하지만, 지금은 그 기능이 이전 버전에서는 작동하지 않는다. 서울은 부산입니다학과를 졸업하고 서울에 입사한 이력이 있는 강사는 김용수(金容洙), 김용수(金容洙), 김용원(金容元)이에요

[ Prompt: 255.9 t/s | Generation: 158.8 t/s ]

> python 예제
요

[ Prompt: 460.2 t/s | Generation: 113.4 t/s ]

> golang 예제
많으셨어요?

[ Prompt: 234.1 t/s | Generation: 164.8 t/s ]

> 많겠냐구요
많으셨어요

[ Prompt: 358.2 t/s | Generation: 162.3 t/s ]

> 아니이~
안녕하세요~ 너무 편하게 물어보세요

[ Prompt: 244.9 t/s | Generation: 138.8 t/s ]

> 안편하다고
보통 대화할 때가 되자요

[ Prompt: 349.4 t/s | Generation: 163.6 t/s ]

> 전화줘
처음 대화하는 시간을 줄이거나 대화하는 시간을 줄인 편하게 물어보세요

[ Prompt: 317.0 t/s | Generation: 152.1 t/s ]

> 안편한데?
질문에 대답하지 마세요

[ Prompt: 544.9 t/s | Generation: 149.4 t/s ]

> 야이
더 쉽게 대화할 수 있어요, 더 잘 보셨나요?

[ Prompt: 267.9 t/s | Generation: 140.7 t/s ]

대충 9 epoch / 23118 스텝 학습

  ┌───────────────┬─────────────────────────────────────────────────────┐
  │               │                                                               │
  ├───────────────┼─────────────────────────────────────────────────────┤
  │ base 사전학습 │ 23,118스텝, 실제 9.42 epoch, val loss 3.4899(ppl 32.8)        │
  ├───────────────┼─────────────────────────────────────────────────────┤
  │ SFT           │ 176스텝(206개 예시), 최고 val loss 2.5465                     │
  ├───────────────┼─────────────────────────────────────────────────────┤
  │ GGUF          │ base/it × f32/q4_0 총 4개 파일 모두 3_small/gguf/에 준비 완료 │
  └───────────────┴─────────────────────────────────────────────────────┘

  ┌─────────────────────┬──────────────────────┬──────────────────────────────┐
  │       데이터        │         용도         │                    규모                     │
  ├─────────────────────┼──────────────────────┼──────────────────────────────┤
  │ 한국어 위키백과     │ 사전학습 본문        │ 문서 676,120개, 문단 316만개, 글자 5.58억자 │
  ├─────────────────────┼──────────────────────┼──────────────────────────────┤
  │ 국립국어원 사전 3종 │ 토크나이저 어휘 시드 │ 뜻풀이 101만 줄, 표제어 32.6만 줄           │
  ├─────────────────────┼──────────────────────┼──────────────────────────────┤
  │ 직접 작성 대화 예시 │ SFT(대화 형식)       │ 206개                                       │
  └─────────────────────┴──────────────────────┴──────────────────────────────┘

  ┌───────┬─────────┬───────────┬───────────────────────┐
  │       │  문서   │   문단    │         토큰          │
  ├───────┼─────────┼───────────┼───────────────────────┤
  │ train │ 672,731 │ 3,144,350 │ 243,428,930 (2.434억) │
  ├───────┼─────────┼───────────┼───────────────────────┤
  │ val   │ 3,389   │ 16,372    │ 1,250,690             │
  └───────┴─────────┴───────────┴───────────────────────┘

  v1 최종 학습량: 약 22.9억 토큰 (2.434억 × 9.42 epoch)

  모델

  - 42.3M 파라미터 (8층, 512차원), fp32
  - 임베딩(16.6M)이 전체의 약 39% 차지

  SFT

  - 206개 대화 예시(직접 작성, 라이선스 깨끗함), 최종 176스텝(약 8 epoch) 학습

  요약하면: 실제 원본 텍스트는 위키 5.58억자를 2.434억 토큰으로 압축해서 갖고 있고, 학습 때는 그걸 약 9.4번 반복해서 총 22.9억 토큰어치를
  모델에 흘려보낸 셈입니다.

-rw-rw-r-- 1 minimonk minimonk 163M  9월 23 14:50 base42m-f32.gguf
-rw-rw-r-- 1 minimonk minimonk 163M  9월 23 14:51 base42m-it-f32.gguf
-rw-rw-r-- 1 minimonk minimonk 28M  9월 23 14:51 base42m-it-q4_0.gguf
-rw-rw-r-- 1 minimonk minimonk 28M  9월 23 14:50 base42m-q4_0.gguf

저작자표시 (새창열림)

'모종의 음모 > ai 프로그램' 카테고리의 다른 글

diffusion gemma 실행해보기 (0)	2026.09.24
diffusion gemma (0)	2026.09.23
wikimatrix (0)	2026.09.22
직접 LLM 만들...기? (0)	2026.09.22
llama-tokenize (0)	2026.09.22

Posted by 구차니

max7219 8-Digit LED Display Drivers

아두이노 예제밖에 안보이냐..

[링크 : https://mobilitykids.tistory.com/entry/MAX7219-LED-매트릭스-디스플레이Arduino-튜토리얼-포함]

[링크 : https://m.blog.naver.com/eduino/222651170585]

spi 같긴한데.. 원래는 7seg 용으로 개발되었으나 dot matrix에서도 쓰는건가?

Features
● 10MHz Serial Interface
● Individual LED Segment Control
● Decode/No-Decode Digit Selection
● 150μA Low-Power Shutdown (Data Retained)
● Digital and Analog Brightness Control
● Display Blanked on Power-Up
● Drive Common-Cathode LED Display
● Slew-Rate Limited Segment Drivers for Lower EMI (MAX7221)
● SPI, QSPI, MICROWIRE Serial Interface (MAX7221)

[링크 : https://www.analog.com/media/en/technical-documentation/data-sheets/max7219-max7221.pdf]

+

선도 연결방법 알려달라고 하고

어떻게 어떻게 해서 claude가 다 해줘버리네. 인간은 멀 취미로 해야하지?

ai가 해준걸 취미라고 해도 될까? 고민되네..

/*
* max7219.h
*
* MAX7219 8x8 도트매트릭스(데이지체인) 구동 + 텍스트 스크롤 드라이버.
* SPI(하드웨어)로 DIN/CLK를 보내고, CS(LOAD)는 별도 GPIO로 직접 제어한다.
*/
#ifndef __MAX7219_H
#define __MAX7219_H

#ifdef __cplusplus
extern "C" {
#endif

#include "stm32f1xx_hal.h"

/* 체인에 연결된 8x8 모듈 개수. 모듈 1개만 있으면 1로 바꿀 것. */
#define MAX7219_NUM_DEVICES   4

/* 하드웨어 배선에 따라 좌우/모듈 순서가 반대로 보이면 아래 두 값을 바꿔서 조정 */
#define MAX7219_REVERSE_CHAIN 1   /* 텍스트가 모듈 순서상 반대로 표시되면 1 (실측: 4 3 2 1로 나와서 1로 설정) */
#define MAX7219_MIRROR_COLS   0   /* 글자가 좌우 반전되어 보이면 1 */

void MAX7219_Init(SPI_HandleTypeDef *hspi);
void MAX7219_Clear(void);
void MAX7219_SetIntensity(uint8_t level /* 0x0~0xF */);

/* 표시할 문자열을 설정(내부 컬럼 버퍼에 렌더링해 둠). ASCII만 지원(공백,0-9,A-Z,.,!,-) */
void MAX7219_SetText(const char *text);

/* 한 프레임(1컬럼) 스크롤 진행 + 출력. 텍스트가 한 바퀴 돌면 1을 반환(그 외 0). */
uint8_t MAX7219_ScrollStep(void);

/* 진단용: 스크롤 없이 모듈(0번부터)마다 고정 숫자 1,2,3,4...를 동시에 정적으로 표시.
* 다중 디바이스 주소 지정이 정상인지 확인할 때만 임시로 사용. */
void MAX7219_ShowDeviceIndexTest(void);

#ifdef __cplusplus
}
#endif

#endif /* __MAX7219_H */

/*
* max7219.c
*
* MAX7219 8x8 도트매트릭스 체인 구동 + 좌->우 텍스트 스크롤.
*
* 배선(하드웨어 SPI1 사용):
*   MAX7219 DIN  -> PA7 (SPI1_MOSI)
*   MAX7219 CLK  -> PA5 (SPI1_SCK)
*   MAX7219 CS   -> PA4 (일반 GPIO, 소프트웨어 제어)
*   MAX7219 VCC  -> 5V,  GND -> GND
*
* 레지스터 모델: MAX7219는 디코드모드 OFF 상태에서 DIGIT0~7 레지스터 각각이
* 매트릭스의 "행(row)"이며, 그 8bit가 그 행의 8개 열(column)을 나타낸다.
* 따라서 프레임을 만들 때는 (행, 각 모듈별 8bit) 형태로 SPI에 실어야 한다.
*/
#include "main.h"
#include "max7219.h"
#include <string.h>
#include <ctype.h>

#define MAX7219_REG_NOOP        0x00
#define MAX7219_REG_DIGIT0      0x01
#define MAX7219_REG_DECODEMODE  0x09
#define MAX7219_REG_INTENSITY   0x0A
#define MAX7219_REG_SCANLIMIT   0x0B
#define MAX7219_REG_SHUTDOWN    0x0C
#define MAX7219_REG_DISPLAYTEST 0x0F

#define MAX7219_TOTAL_COLS      (MAX7219_NUM_DEVICES * 8)

/* 표시 가능한 최대 문자열 길이(글자당 6컬럼 = 5폭+1간격) + 앞뒤 여백 */
#define MAX7219_TEXT_MAX_CHARS  64
#define MAX7219_COLBUF_SIZE     (MAX7219_TOTAL_COLS * 2 + MAX7219_TEXT_MAX_CHARS * 6)

static SPI_HandleTypeDef *s_hspi;
static uint8_t  s_colbuf[MAX7219_COLBUF_SIZE];
static uint32_t s_colbuf_len;
static uint32_t s_scroll_offset;

/* ---- 5x7 폰트: 행(row) 단위 원본 표현. 각 바이트의 bit4..bit0 = 좌->우 5개 컬럼 ---- */
typedef struct { char ch; uint8_t rows[7]; } Glyph5x7;

static const Glyph5x7 s_font[] = {
  {' ', {0b00000,0b00000,0b00000,0b00000,0b00000,0b00000,0b00000}},
  {'0', {0b01110,0b10001,0b10011,0b10101,0b11001,0b10001,0b01110}},
  {'1', {0b00100,0b01100,0b00100,0b00100,0b00100,0b00100,0b01110}},
  {'2', {0b01110,0b10001,0b00001,0b00010,0b00100,0b01000,0b11111}},
  {'3', {0b11111,0b00010,0b00100,0b00010,0b00001,0b10001,0b01110}},
  {'4', {0b00010,0b00110,0b01010,0b10010,0b11111,0b00010,0b00010}},
  {'5', {0b11111,0b10000,0b11110,0b00001,0b00001,0b10001,0b01110}},
  {'6', {0b00110,0b01000,0b10000,0b11110,0b10001,0b10001,0b01110}},
  {'7', {0b11111,0b00001,0b00010,0b00100,0b01000,0b01000,0b01000}},
  {'8', {0b01110,0b10001,0b10001,0b01110,0b10001,0b10001,0b01110}},
  {'9', {0b01110,0b10001,0b10001,0b01111,0b00001,0b00010,0b01100}},
  {'A', {0b01110,0b10001,0b10001,0b11111,0b10001,0b10001,0b10001}},
  {'B', {0b11110,0b10001,0b10001,0b11110,0b10001,0b10001,0b11110}},
  {'C', {0b01111,0b10000,0b10000,0b10000,0b10000,0b10000,0b01111}},
  {'D', {0b11110,0b10001,0b10001,0b10001,0b10001,0b10001,0b11110}},
  {'E', {0b11111,0b10000,0b10000,0b11110,0b10000,0b10000,0b11111}},
  {'F', {0b11111,0b10000,0b10000,0b11110,0b10000,0b10000,0b10000}},
  {'G', {0b01111,0b10000,0b10000,0b10011,0b10001,0b10001,0b01111}},
  {'H', {0b10001,0b10001,0b10001,0b11111,0b10001,0b10001,0b10001}},
  {'I', {0b01110,0b00100,0b00100,0b00100,0b00100,0b00100,0b01110}},
  {'J', {0b00111,0b00010,0b00010,0b00010,0b00010,0b10010,0b01100}},
  {'K', {0b10001,0b10010,0b10100,0b11000,0b10100,0b10010,0b10001}},
  {'L', {0b10000,0b10000,0b10000,0b10000,0b10000,0b10000,0b11111}},
  {'M', {0b10001,0b11011,0b10101,0b10101,0b10001,0b10001,0b10001}},
  {'N', {0b10001,0b11001,0b10101,0b10101,0b10011,0b10001,0b10001}},
  {'O', {0b01110,0b10001,0b10001,0b10001,0b10001,0b10001,0b01110}},
  {'P', {0b11110,0b10001,0b10001,0b11110,0b10000,0b10000,0b10000}},
  {'Q', {0b01110,0b10001,0b10001,0b10001,0b10101,0b10010,0b01101}},
  {'R', {0b11110,0b10001,0b10001,0b11110,0b10100,0b10010,0b10001}},
  {'S', {0b01111,0b10000,0b10000,0b01110,0b00001,0b00001,0b11110}},
  {'T', {0b11111,0b00100,0b00100,0b00100,0b00100,0b00100,0b00100}},
  {'U', {0b10001,0b10001,0b10001,0b10001,0b10001,0b10001,0b01110}},
  {'V', {0b10001,0b10001,0b10001,0b10001,0b10001,0b01010,0b00100}},
  {'W', {0b10001,0b10001,0b10001,0b10101,0b10101,0b10101,0b01010}},
  {'X', {0b10001,0b10001,0b01010,0b00100,0b01010,0b10001,0b10001}},
  {'Y', {0b10001,0b10001,0b01010,0b00100,0b00100,0b00100,0b00100}},
  {'Z', {0b11111,0b00001,0b00010,0b00100,0b01000,0b10000,0b11111}},
  {'.', {0b00000,0b00000,0b00000,0b00000,0b00000,0b01100,0b01100}},
  {'!', {0b00100,0b00100,0b00100,0b00100,0b00100,0b00000,0b00100}},
  {'-', {0b00000,0b00000,0b00000,0b11111,0b00000,0b00000,0b00000}},
};
#define FONT_COUNT (sizeof(s_font) / sizeof(s_font[0]))

static const Glyph5x7 *font_lookup(char c)
{
  if (c >= 'a' && c <= 'z') c = (char)(c - 'a' + 'A'); /* 소문자 -> 대문자 */
  for (uint32_t i = 0; i < FONT_COUNT; i++)
  {
    if (s_font[i].ch == c) return &s_font[i];
  }
  return &s_font[0]; /* 지원 안 하는 문자는 공백으로 */
}

/* row(0=top..6=bottom) 문자열 비트에서 col(0=left..4)에 해당하는 컬럼 바이트를 뽑아
* bit r 위치에 채운다 (bit0=top row ... bit6=bottom row). */
static uint8_t glyph_col_byte(const Glyph5x7 *g, uint8_t col)
{
  uint8_t val = 0;
  for (uint8_t r = 0; r < 7; r++)
  {
    uint8_t bit = (g->rows[r] >> (4 - col)) & 0x01U;
    val |= (uint8_t)(bit << r);
  }
  return val;
}

/* ---- 저수준 SPI 전송 ---- */

static inline void cs_low(void)  { HAL_GPIO_WritePin(MAX7219_CS_PORT, MAX7219_CS_PIN, GPIO_PIN_RESET); }
static inline void cs_high(void) { HAL_GPIO_WritePin(MAX7219_CS_PORT, MAX7219_CS_PIN, GPIO_PIN_SET); }

/* 체인의 모든 모듈에 동일한 (reg,data)를 전송 (초기화/공통 설정용) */
static void max7219_write_all(uint8_t reg, uint8_t data)
{
  uint8_t buf[2] = { reg, data };
  cs_low();
  for (uint32_t i = 0; i < MAX7219_NUM_DEVICES; i++)
  {
    HAL_SPI_Transmit(s_hspi, buf, 2, HAL_MAX_DELAY);
  }
  cs_high();
}

/* 한 행(row 0~7)에 대해 모듈별 8bit 데이터를 체인에 맞는 순서로 전송.
* dev_data[0] = MCU와 가장 가까운(첫번째) 모듈. SPI 시프트 특성상 체인에서는
* 가장 먼 모듈 데이터부터 먼저 밀어넣어야 CS 상승 시 올바르게 래치된다. */
static void max7219_send_row(uint8_t row, const uint8_t *dev_data)
{
  uint8_t reg = (uint8_t)(MAX7219_REG_DIGIT0 + row);
  cs_low();
#if MAX7219_REVERSE_CHAIN
  for (int32_t dev = 0; dev < MAX7219_NUM_DEVICES; dev++)
#else
  for (int32_t dev = MAX7219_NUM_DEVICES - 1; dev >= 0; dev--)
#endif
  {
    uint8_t buf[2] = { reg, dev_data[dev] };
    HAL_SPI_Transmit(s_hspi, buf, 2, HAL_MAX_DELAY);
  }
  cs_high();
}

void MAX7219_Init(SPI_HandleTypeDef *hspi)
{
  s_hspi = hspi;

  max7219_write_all(MAX7219_REG_DISPLAYTEST, 0x00);
  max7219_write_all(MAX7219_REG_SCANLIMIT, 0x07);   /* 8자리(행) 전부 사용 */
  max7219_write_all(MAX7219_REG_DECODEMODE, 0x00);  /* BCD 디코드 없음: 매트릭스 모드 */
  max7219_write_all(MAX7219_REG_INTENSITY, 0x04);   /* 밝기 0x0~0xF */
  max7219_write_all(MAX7219_REG_SHUTDOWN, 0x01);    /* 0=shutdown, 1=normal operation */

  MAX7219_Clear();
}

void MAX7219_SetIntensity(uint8_t level)
{
  max7219_write_all(MAX7219_REG_INTENSITY, (uint8_t)(level & 0x0FU));
}

void MAX7219_Clear(void)
{
  uint8_t zeros[MAX7219_NUM_DEVICES];
  memset(zeros, 0, sizeof(zeros));
  for (uint8_t row = 0; row < 8; row++)
  {
    max7219_send_row(row, zeros);
  }
}

void MAX7219_SetText(const char *text)
{
  uint32_t len = 0;

  /* 화면 밖에서 스크롤 들어오도록 앞쪽 여백(패널 폭만큼 공백) */
  for (uint32_t i = 0; i < MAX7219_TOTAL_COLS && len < MAX7219_COLBUF_SIZE; i++)
  {
    s_colbuf[len++] = 0x00;
  }

  for (const char *p = text; *p != '\0' && len + 6 <= MAX7219_COLBUF_SIZE; p++)
  {
    const Glyph5x7 *g = font_lookup(*p);
    for (uint8_t c = 0; c < 5; c++)
    {
      s_colbuf[len++] = glyph_col_byte(g, c);
    }
    s_colbuf[len++] = 0x00; /* 글자 사이 1컬럼 간격 */
  }

  /* 화면 밖으로 완전히 빠져나가도록 뒤쪽 여백 */
  for (uint32_t i = 0; i < MAX7219_TOTAL_COLS && len < MAX7219_COLBUF_SIZE; i++)
  {
    s_colbuf[len++] = 0x00;
  }

  s_colbuf_len = len;
  s_scroll_offset = 0;
}

static void render_frame(void)
{
  uint8_t dev_row[MAX7219_NUM_DEVICES];

  if (s_colbuf_len == 0) return;

  for (uint8_t row = 0; row < 8; row++)
  {
    for (uint32_t dev = 0; dev < MAX7219_NUM_DEVICES; dev++)
    {
      uint8_t val = 0;
      for (uint8_t c = 0; c < 8; c++)
      {
        uint32_t col_index = (s_scroll_offset + dev * 8U + c) % s_colbuf_len;
        uint8_t bit = (s_colbuf[col_index] >> row) & 0x01U;
#if MAX7219_MIRROR_COLS
        val |= (uint8_t)(bit << c);
#else
        val |= (uint8_t)(bit << (7 - c));
#endif
      }
      dev_row[dev] = val;
    }
    max7219_send_row(row, dev_row);
  }
}

uint8_t MAX7219_ScrollStep(void)
{
  render_frame();

  if (s_colbuf_len == 0) return 1;

  s_scroll_offset++;
  if (s_scroll_offset >= s_colbuf_len)
  {
    s_scroll_offset = 0;
    return 1;
  }
  return 0;
}

/* 각 모듈에 스크롤 없이 고정 숫자를 동시에 정적으로 표시(디버그용).
* dev=0 -> "1", dev=1 -> "2" ... 모듈이 서로 다른 내용을 "동시에" 보여주는지,
* 그리고 계속 켜진 상태로 유지되는지(깜빡임/순차갱신 없는지) 확인하는 용도. */
void MAX7219_ShowDeviceIndexTest(void)
{
  uint8_t dev_row[MAX7219_NUM_DEVICES];
  const Glyph5x7 *glyphs[MAX7219_NUM_DEVICES];

  for (uint32_t dev = 0; dev < MAX7219_NUM_DEVICES; dev++)
  {
    char digit = (char)('1' + (dev % 9));
    glyphs[dev] = font_lookup(digit);
  }

  for (uint8_t row = 0; row < 8; row++)
  {
    for (uint32_t dev = 0; dev < MAX7219_NUM_DEVICES; dev++)
    {
      uint8_t val = 0;
      if (row < 7) /* 글꼴은 7행(0~6)까지만 있음, 8번째 행은 항상 빈 줄 */
      {
        for (uint8_t c = 0; c < 5; c++) /* 글꼴 폭 5, 나머지 3컬럼은 여백(0) */
        {
          uint8_t bit = (glyphs[dev]->rows[row] >> (4 - c)) & 0x01U;
#if MAX7219_MIRROR_COLS
          val |= (uint8_t)(bit << (2 + c));
#else
          val |= (uint8_t)(bit << (7 - c));
#endif
        }
      }
      dev_row[dev] = val;
    }
    max7219_send_row(row, dev_row);
  }
}

저작자표시 (새창열림)

'embeded > Cortex-M3 STM' 카테고리의 다른 글

ws2812b vs hub75 (0)	2026.09.23
네오픽셀 ws2812b-64 작동 (0)	2026.09.23
stm32 gpio hi-z (0)	2026.09.10
stm32f103c9t6(bluepill) + keypad (0)	2026.09.08
stm32 usb composite / report id (0)	2026.09.04

Posted by 구차니

구차니의 잡동사니 모음

'2026/09'에 해당되는 글 88건

kicad 공식 사용법

'프로그램 사용 > kicad' 카테고리의 다른 글

kicad 발음

'프로그램 사용 > kicad' 카테고리의 다른 글

스타크래프트 ai의 시대

'게임 > 스타크래프트 시리즈' 카테고리의 다른 글

llm 초기값

'모종의 음모 > ai 프로그램' 카테고리의 다른 글

피곤

'개소리 왈왈 > 육아관련 주저리' 카테고리의 다른 글

흐음..

'개소리 왈왈 > 인공지능' 카테고리의 다른 글

diffusion gemma 실행해보기

'모종의 음모 > ai 프로그램' 카테고리의 다른 글

diffusion gemma

'모종의 음모 > ai 프로그램' 카테고리의 다른 글

직접만든 LLM .. part 2

'모종의 음모 > ai 프로그램' 카테고리의 다른 글

max7219 8-Digit LED Display Drivers

'embeded > Cortex-M3 STM' 카테고리의 다른 글

카테고리

공지사항

태그목록

최근에 올라온 글

최근에 달린 댓글

티스토리툴바