
Qwen3.8-27Bã¯RTX 4090 24GBã§åãã®ãïŒ LM StudioããUnslothãŸã§ãããŒã«ã«LLMã®éçã宿©æ€èšŒ

ããã«ã¡ã¯ãSTUDIO55æè¡çµ±æ¬ã®å ¥æ±ã§ãã
Metaã® Muse Glimmer-30B ãå ¬éãããããã 4æ¥åŸãAlibabaã® Qwenãã Qwen3.8-27B ãå ¬éãããŸããã
ã©ã¡ããåãã30Bã¯ã©ã¹ã®ããŒã«ã«LLMãã§ãããããšãŒãžã§ã³ãç³»AIã¢ãã«ãã§ãã
ãã®å€ãMetaãš Alibabaãšãã2瀟ãããããŒã«ã«ã§åäœãã倧åã¢ãã«ããããã æ°æ¥ã®éã«çžæ¬¡ãã§ç»å Žããããšã«ãªããŸãã
ãããŠçŸåšãäžã®äžã§ã¯ãGPT-6 Astraãã倧ããªè©±é¡ã«ãªã£ãŠããŸãã
Astra ã¯ãæããæåº§ããæå³ããèšèã§ãã

8æ1æ¥ã«OpenAI ãæ¬¡æäž»èŠã¢ãã«ãå ¬è¡šãã9æ3æ¥ã«æ£åŒãªãªãŒã¹ãå ±éæ©é¢åãã®ããªãŒãã£ã³ã°ã§ã¯ããAGIæä»£ãžããããããšããèšèã§ç· ããããããã»ã©ããããŸã§ã®AI ãšã¯äžç·ãç»ãæ©èœãæ«é²ãããŠããŸãã
ãã§ã«ã³ãã¥ããã£ã§ã¯ããAGI ã¯ãã€ããŒã«ã«ã§äœ¿ããããã«ãªãã®ãããšãã£ã話é¡ãŸã§äžãã£ãŠãããAI ã®é²åã«å¯ŸããæåŸ
ã¯ãããªãåã®ããã«ãªã£ãŠããŸãã
ãã®èšäºã§ã¯ããããŸã§æ€èšŒããŠãã Muse Glimmerã«ç¶ããQwen3.8-27B ãåãäžããŸãã
ãããŠã2ã€ã®ãšãŒãžã§ã³ãç³»AIã¢ãã«ãå®éã«æ€èšŒããããšã§ãåã«ãã©ã¡ãã®ã¢ãã«ãåªããŠãããããæ¯èŒããã®ã§ã¯ãªããAI ãã©ã®ããã«ãã·ãŒã³ãåŠçããå€éšã®ããŒã«ãããŒã¿ãšã€ãªãã£ãŠããã®ããšããããã®çµè·¯ãã®ãã®ã«ç®ãåããŠãããŸãã
ããŒã«ã«AI ãå®éã«æ±ã£ãŠã¿ããšãã¯ã©ãŠãAI ã䜿ã£ãŠããã ãã§ã¯èŠãã«ãããAI ãšãã·ãŒã³ã®é¢ä¿ãåŠçã®ä»çµã¿ãèŠããŠããŸãã
ãAIãªããªãã§ãã§ããããšæããã»ã©é«æ§èœã«ãªã£ãçŸåšã§ããå®éã«ã¯èšç®è³æºãã¡ã¢ãªãGPUãã³ã³ããã¹ããããŒã«ããããŠåŠçæéãšãã£ãã³ã¹ããååšããŸãã
ã€ãŸããAI ã¯æ±ºããŠç¡éã«äœã§ãåŠçã§ããããã§ã¯ãããŸããã
ãã®ä»çµã¿ãçè§£ããããšã¯ãããŒã«ã«LLM ã䜿ãããã ãã§ã¯ãªããããããAI ãã®ãã®ãã©ããžåããããšããŠããã®ããèããããã§ãéèŠã§ãã
ãšã³ãžãã¢äžå¿ã®ã³ãã¥ããã£ã§é£ã³äº€ã£ãŠããããŸããŸãªAI ã®è©±é¡ããããããä»çµã¿ãç¥ã£ãŠããããšã§ãåãªããAIã®ããã話ãã§ã¯ãªãããã®è£åŽã§äœãèµ·ããŠããã®ããšããèŠç¹ããèªã¿è§£ããããã«ãªãã¯ãã§ãã
ð€Qwen3.8-27B
ãã®å€ãAIã®é²åãããããŸã§ä»¥äžã«éãã¹ããŒãã§é²ãã§ããããšããæ¹ããŠå®æãããããåºæ¥äºãç¶ããŸããã

ãšãã«ãQwen3.8-27Bãã¯ãå ¬éããããã2æ¥ã§ã³ãã¥ããã£ããçŽ500çš®é¡ã®éååçãå ¬éããããªã©ãéåžžã«å€§ããªçãäžãããèŠããŠããŸãã
ãã®ãããªæµãã¯ãããæå³å¿ ç¶ã§ãQwenã·ãªãŒãºã¯åŸæ¥ããã³ãã¥ããã£ã§ã®æŽ»çšãåæãšãããšã³ã·ã¹ãã ã圢æãããŠããŸããããã«å ããä»åã®Â Qwen3.8-27BïŒå ¬åŒãªããžããªïŒã¯ safetensors圢åŒã§18åå²ãåèš55.56GBãšããéåžžã«å€§ããªã¢ãã«ã§ãå®è³ªçã«ããŒã¿ã»ã³ã¿ãŒåãGPU ãå¿ èŠãšãããã¯ã©ã¹ã§ãã
ãã®ãããå€ãã®ãŠãŒã¶ãŒã¯ç©æ¥µçã«éååçãå©çšããããCPUã RAMãžã®ãªãããŒããçµã¿åãããéçšã§çãäžãã£ãŠããŸãã

ã¡ãªã¿ã«ãéååçã§ã¯Â Unslothã® Qwen3.8-27B NVFP4 ãéåžžã«é«ã人æ°ãéããŠããŸãã2026幎8æ25æ¥çŸåšã630äžããŠã³ããŒãã2.6Kã®ããããããèšé²ããŠããŸãã
ãã ããããã«ã¯æ³šæãå¿
èŠã§ãã
Unslothã® NVFP4ãªããžããªã«ã¯ "Why no NVFP4 GGUF?" ãšããDiscussionãããããã«ããã®ã¢ãã«ã¯Â GGUFçã§ã¯ãªããSafetensorsããŒã¹ã® NVFP4ã¢ãã«ã§ãã

NVFP4 ã¯ãNVIDIAã Blackwelläžä»£ã§å°å ¥ãã 4bitã®æµ®åå°æ°ç¹ãã©ãŒãããã§ãäœãããåã«ããã¡ã¢ãªäœ¿çšéã®åæžãšã粟床äœäžã®æå¶ãçã£ããã®ã§ããNVFP4 ããã€ãã£ãã«é«éåŠçã§ããããŒããŠã§ã¢æ©èœã¯ãBlackwelläžä»£ïŒRTX 50ã·ãªãŒãºïŒããæèŒãããŠããŸãã
ç§ã䜿çšããŠããGPUã¯Â RTX 4090ã§ãAda Lovelaceäžä»£Â ã§ãããã® RTX 4090㯠LLMã®ããã«çšæããæ©æã§ã¯ãªããããšããšCGå¶äœåãã®GPUã¬ã³ããªã³ã°ã«äœ¿çšããŠããæ¥åçšãã·ã³ã§ãã
ãã®ãããBlackwelläžä»£åãã®NVFP4ããŒããŠã§ã¢ã¢ã¯ã»ã©ã¬ãŒã·ã§ã³ã¯å©çšã§ããŸããã
ã€ãŸããæ°ãã« LLMå°çšã® GPUãå°å ¥ããããšãªããæ¥åžžæ¥åã§äœ¿çšããŠãã RTX 4090ããã®ãŸãŸæŽ»çšã§ããã®ã ââ ããããä»åã®æ€èšŒããŒãã® 1ã€ã§ããããŸãã
ð¬LM Studio ã§æ€èšŒããçµæããŸãäŒããŸã
ååã«ç¶ããŠããŸã LM Studioã§ Qwen3.8 ãåããŸãã
LM Studio ã¢ããªå
ã§ãShow filesããæŒããšããã®ã¢ãã«ã§çšæãããŠãããã¡ã€ã«ïŒéååããªã¢ã³ãïŒã®äžèЧãå±éãããã®ã§ãããããéžæããŸãã

LM Studioã§ã¯ã䜿çšããŠããããŒããŠã§ã¢ã«å¿ããŠã察å¿ããã¢ãã«ãæšå¥šãããéååããªã¢ã³ãã衚瀺ããããã®ç°å¢ã§çŸå®çã«åãããªããã®ã¯ãããããéžæè¢ãšããŠåºãŠããŸããã

ã芧ã®ããã«ãå·Šã®Staff Pick ã®äžçªäžã«è¡šç€ºããããFull GPU Offload Possible ã®ç·ãããžãä»ããŠããŸãã
ããã¯ãç§ã䜿çšãããã·ãŒã³ "4090ã® 24GBã«å
šã¬ã€ã€ãŒãèŒã" ãšããæå³ã§ãã
ä»åã®æ€èšŒã«äœ¿ã£ãã¢ãã«ã¯ããã®ãããã«è¡šç€ºããããQwen3.8 27B Q4_L_MïŒ18.16 GBïŒãGGUF ã§ãã
ðMCPãšããã¥ã¡ã³ãããŒã¿ã®è§£æãã¹ã
çµè«ãããäŒãããŸãã
MCPãã¹ãã§ããDãã©ã€ãå
ã®ããŒã¿ãæŽçãããããæç€ºãããšãããMuse Glimmer ã®å Žåã«ã¯ããã©ã®ããã«æŽçããã°ããããã確èªããŠããŸãã

äžæ¹ãQwen3.8 ã§ã¯è¿œå ã®ç¢ºèªãè¡ãããèªèº«ã®å€æã§ãã©ã«ããäœæããŠæŽçãé²ããŸããã

ãã®ãããªå·®ããQwen3.8 ããšãŒãžã§ã³ãçšéãæç€ºçã«åŒ·åããã¢ãã«ã ãšæèšã§ãããã©ããã¯åãããŸãããããšã«ãããåæã«ãã£ãŠãããããšãã£ãå°è±¡ããã£ãã®ã¯ç¢ºãã§ãã
Developer Role ãžã®å¯Ÿå¿ãããŒã«åŒã³åºãã®æ¹åãªã©ããéäžã§æ¢ãŸãããèªåŸçã«åŠçãé²ããæ¹åã®èšèšææ³ãšæŽåããŠããŸããCodex ã®ãããªããŒã«ç°å¢ã§ã®å©çšãæèããã¢ãã«ãšèšããã®ã§ãããã
ãã®åŸåã¯ãå³é¢PDFã®ãã¹ãã§ãå ±éããååã®ãã¹ããšåãå³é¢PDFãèªãŸãããšãããQwen ã¯èããããŠéäžã§åæ¢ããŠããŸããŸããã
6ããŒãžãã PDFã§ããããšãã®éäžã§ã³ã³ããã¹ãé·ã䜿ãåã£ãŠããŸã£ãã®ã§ãã

Glimmer ã¯ãæž¡ããã PDFããã®ãŸãŸåŠçããŠåçãè¿ããŸããããQwen3.8 ã¯ãããé«ãè§£å床ã§åŠçããçµè·¯ãéžæããããšããçµæããã®å®è¡ãç°å¢ã®å¶çŽã«åãŸããªãã£ãããã§ãã
èŠããã«ãèœåã®éãã§ã¯ãªããäž¡ã¢ãã«ãç°ãªã åŠççµè·¯Â ãéžæãããšããããšã§ãã
Qwen ã¯åŠçã®éäžã§ãããé©åãšæãããæ¹æ³ã èªãéžæããããšãããç¶æ³ã倿ããå¿ èŠã§ããã°åŠçæ¹æ³ã倿Žããããã®ãåŠçæ¹æ³ãèªåŸçã«éžæãããæ§è³ªããå³é¢èªè§£ã§è£ç®ã«åºããšè§£éã§ããŸãã
ããã¯ãRedditã§ãèŠããã
ãxhighã¯èããããŠ4åã®æéããããŠå€±æããã
ãšããå ±åãšããæ§é çãªå
±éç¹ãæããŸãã
ã¢ãã«ãããéãæšè«ã»åŠççµè·¯ãéžæããã»ã©ãå®è¡æéãšãªãœãŒã¹æ¶è²»ã¯å¢å ããŸãããã®çµæãå®è¡ç°å¢ã®å¶çŽã«éããåŠçãå®äºã§ããªããªããŸãã
ãã®ããšã¯ãVRAM ãåçŽã«å¢ããã°è§£æ±ºããåé¡ãšã¯å°ãéããŸãã
ããšãã° 8GBå€ã GPUã§ããã°æ¹åããå¯èœæ§ã¯ãããŸãããããã2000Ã1414ãã¯ã»ã«ã®ç»åã 6æåŠçã§ããä¿èšŒãããããã§ã¯ãããŸãããä»®ã«ãã®åŠçãå¯èœã«ãªã£ããšããŠããã¢ãã«ãããã«èªåã®èãã§é«ãè§£å床ãéžæããã°ãåãåé¡ãåã³èµ·ããå¯èœæ§ããããŸãã
äžæ¹ã§ãã³ãã¥ããã£ã«ã¯ãåã 24GBã® GPUã§ 10äžããŒã¯ã³èŠæš¡ã®ã³ã³ããã¹ãé·ã確ä¿ããŠéçšããŠããäŸããããŸããæ°ãã GPUã ããå®çŸã§ããã®ã§ã¯ãªããæšè«æã®èšå®ã調æŽããçµæã§ãã
ã€ãŸãåé¡ã¯ãããŒããŠã§ã¢ãã®ãã®ã§ã¯ãªããã¢ãã«ãéžæããåŠçãå®è¡ç°å¢ã®ç¯å²å ãžåãããããããå©çšè ãã©ããŸã§å¶åŸ¡ã§ããã ã«ãããŸãã
ããããLM Studio ã§ã¯ããã®èª¿æŽãèŠã€ãããŸããã§ããã
ðæ€èšŒã¡ã¢
ä»åã®æ€èšŒã§ã¯ãQwen3.8 ããåããŒãžãé«è§£å床ã§è©³ããèªã¿åãããšããåŠççµè·¯ãèªãéžæããPDFã®6ããŒãžãçŽ2000Ã1414px ã§å±éããããšããŸããã
èšå®é ç®ãšããŠçšæãããŠããã®ã¯ Reasoning BudgetïŒ8192ïŒ ã§ãããã¯æèéãããŒã¯ã³æ°ã§å¶éãããã®ã§ããã€ãŸãããã©ã®ããã«èãããããå¶åŸ¡ããã®ã§ã¯ãªãããã©ãã§æã¡åããããæ±ºããèšå®ãšèšããŸãã

Reddit ã§ã¯ãmmproj-offload = false ãæå®ããããžã§ã³åŠçã®äžéšãRAMãžãªãããŒãã㊠GPUã¡ã¢ãªã«äœè£ãæãããæ§æäŸãå ±åãããŠããŸããããããLM Studioã«ã¯åçã®èšå®ã¯èŠåœãããŸããã
GPU Offload ã¯ãªãŒããåãã°æåèšå®ã§ããŸãããå®è¡äžã«å€ãåæã«å€æŽãããæåã確èªãããŸããïŒäŸïŒ65 â 13ïŒã

ãããã®æ€èšŒãããQwen3.8 ãå®å®ããŠæ±ãã«ã¯ãéããããªãœãŒã¹ãå®è¡ç°å¢åŽãã©ãã ãå¹çãã管çã§ãããã課é¡ã«ãªããšããããšãåãããŸããã
å°ãªããšãä»åã®æ€èšŒã§ã¯ãMuse Glimmer ã«ã¯ãå®è¡ç°å¢ã«åãããŠåŠçã調æŽããå®å®ããŠå®äºãããæ±çšæ§ãæããããŸããã
äžæ¹ãQwen3.8ã¯ãããç©æ¥µçã«åŠçãè¡ã代ããã«ãVRAM ãã³ã³ããã¹ãé·ãªã©ãå®è¡ç°å¢ã®åœ±é¿ãåãããããšããå°è±¡ã§ãã
ããã§æ¬¡ã¯ãã¢ãã«ãã®ãã®ã ãã§ãªãããããåãããããŒãã¹ãã®éãã«çç®ãããUnslothãã®ãã¬ãŒã ã¯ãŒã¯ã§ Qwen3.8ãå®è¡ããããšã«ããŸããã
Unsloth ã¯ãåã«ã¢ãã«ãå®è¡ããã ãã§ãªããåŠç¿ã»å€æã»éååã»åçš®ãã©ã¡ãŒã¿ã®èª¿æŽãŸã§æ±ãããããUnslothçã®UD-Q4_K_XLã䜿ãããšã§ãLM Studio ãšã¯ç°ãªãå®è¡æ¡ä»¶ã§æ€èšŒãã§ãããã§ãã
æãããŠãå®è¡ç°å¢ããã现ããã³ã³ãããŒã«ããããšã§ãQwen3.8ã®æåã¯ã©ããŸã§å®å®ããã®ã§ããããã
ð ïžUnsloth ã®æ€èšŒ
æåã«ããUnslothãã«ã€ããŠè§£èª¬ããŠãããŸãã
æè¿ç»å ŽãããUnslothïŒã¢ã³ã¹ããŒã¹ïŒãã¯ãQwen3.8 ãã¯ãããšããææ°LLMã®æšè«ãããã¡ã€ã³ãã¥ãŒãã³ã°ãŸã§ãäžè²«ããŠè¡ããããŒã«ã«å®è¡ç°å¢ãšããŠã倧ããªæ³šç®ãéããŠããŸãã

Unsloth ã®æ¬äœã¯ããã¡ã€ã³ãã¥ãŒãã³ã°çšã® Pythonã©ã€ãã©ãªÂ ã§ãã
ããšããšã¯ãLoRA/QLoRA ã«ããåŠç¿ãé«éãã€çã¡ã¢ãªã§å®è¡ããããšãç®çãšããŠéçºãããHugging Faceã®Â Transformersã TRLãããŒã¹ã«æ§ç¯ãããŠããŸãã
çŸåšã¯ãäž»ã«æ¬¡ã® 3ã€ã®åœ¢æ ã§æäŸãããŠããŸãã
Unsloth Core â ã³ãŒãããŒã¹çïŒpip install unslothïŒããããå ã ã®å§¿
Unsloth Studio â ãã©ãŠã¶UIãhttp://127.0.0.1:8888 ã§åã
Unsloth Desktop â Tauri補ã®ãã¹ã¯ãããã¢ããªãâä»å䜿ããã®
ãªããæšè«æ©èœã®ããã¯ãšã³ãã«ã¯ llama.cpp ãæ¡çšãããŠããŸãã
ã·ã¹ãã ã«ãããGGUF æšè«ãšã³ãžã³ãã®ã»ã¯ã·ã§ã³ã«ã次ã®ããã«ã¯ã£ãããšæžãããŠããŸãã
llama.cpp ã¯çŸåš CUDA ã§åäœããŠããŸãã
ãããŠã«ã¹ã¿ã llama.cpp ãã©ã«ãã®æå®æ¬ã確èªã§ããŸãã

ãã®ãããExtra Argumentsã«ã¯ llama.cppïŒllama-serverïŒãžæž¡ãã³ãã³ãã©ã€ã³åŒæ°ããã®ãŸãŸæå®ã§ããŸãã
å®éã«ãã°ã確èªãããšãèµ·åæã«å®è¡ããã llama-server ã®å®å šãªã³ãã³ãã©ã€ã³ãèšé²ãããŠããŸãã
Starting llama-server: C:\Users\Owner\.unsloth\llama.cpp\build\bin\Release\llama-server.exe
-m ...\Qwen3.8-27B-UD-Q4_K_XL.gguf --port 59336 --parallel 4 --flash-attn on
--no-context-shift -c 8192 -ngl -1 --fit off --metrics ...
--mmproj ...\mmproj-F16.gguf --cache-ram 0 --ctx-checkpoints 0
--rope-scaling yarn --yarn-orig-ctx 32768 --mmproj-offloadããã§æå®ãããŠãã
.unsloth\llama.cpp\build\bin\Release\llama-server.exe
ããå®éã«ã¢ãã«ã®æšè«åŠçãå®è¡ããŠããå®è¡ãã¡ã€ã«ãã®ãã®ã§ãã
ã€ãŸããUnsloth ã®ç»é¢ããã¢ãã«ãèµ·åããŠããããã«èŠããŠãããã®è£åŽã§ã¯ llama.cpp ã® llama-server ãå®éã®æšè«ãšã³ãžã³ãšããŠåäœããŠããããšãããã®ãã°ãã確èªã§ããŸãã
ðŒïžUnslothã§ããŒã«ã«ã§ç»åçæ
ãããããUnslothãã¯ãåçŽã« LM Studioãšåãçš®é¡ã®ããŒãã¹ã§ã¯ãããŸããã
Unslothã¯ãã¢ãã«ã®å®è¡ã ãã§ãªãããã¡ã€ã³ãã¥ãŒãã³ã°ãªã©ãå«ããŠæ±ãçµ±åç㪠AIå®è¡ç°å¢ã§ãããä»ã®ããŒã«ãæšè«åºç€ãšçµã¿åãããŠå©çšããããã®åå°ãšããŠæ©èœããŸãã
ããã«çŸåšã® Unsloth Desktop ã§ã¯ãããŒã«ã«ã§ã®ç»åçæã«ã察å¿ããŠããŸãã
ããããç»åçæã ãã§ã¯ãããŸãããç»åç·šéãåç»çæã«ã察å¿ããŠããããããŸã§ ComfyUI ã Forge ãªã©ãçµã¿åãããŠæ§ç¯ããŠããã¯ãŒã¯ãããŒã®äžéšããUnslothäžã§å®çµãããããšãå¯èœã«ãªã£ãŠããŸãã

ç»åçæïŒFLUXãMiniMax-H3ãZ-Image ãªã©ïŒ
ç»åç·šéïŒInpaintingãOutpaintingãã¢ããã¹ã±ãŒã«ãåç §ç»åã«ããçæãªã©ïŒ
åç»çæïŒWanãLTX ãªã©ïŒ
LoRA ã®å©çšã»åŠç¿
Unsloth ã¯Â ã¢ãã«ãå®è¡ããããã®ãã©ãããã©ãŒã ãšããŠã
DeepSeek-V4 ãéžã¹ã° LLMïŒæç« çæïŒ
FLUX ãéžã¹ã° ç»åçæ
Wan ãéžã¹ã° åç»çæ
ãšããããã«ãèªã¿èŸŒãã¢ãã«ã«ãã£ãŠæ©èœãå€ãããŸãã
åãªããLLMãåãããããã®ããŒãã¹ãããUnslothã¯ãããŒã«ã«AIãæ±ãããã®ãç·åçãªå®è¡ç°å¢ããžãšããã®åœ¹å²ãåºãã€ã€ãããšèšããŸãã
ðŠQwen3.8-27B-GGUF ã®ã€ã³ã¹ããŒã«
ã¢ãã«ã®ã€ã³ã¹ããŒã«ã¯ããã£ããç»é¢äžéšã«ãããSelect modelãããè¡ãããšãã§ããŸãã
LM Studioãšã¯ç°ãªãã該åœãã·ãŒã³ã§å¯Ÿå¿ãé£ããã¢ãã«ãäžèЧã«è¡šç€ºãããŸãã

äžèšç»åã§ãDeepSeek-V4-FlashïŒ284.3BïŒãš DeepSeek-V4-ProïŒ1573.0BïŒã«ãèµ€ã ãOOMãããŒã¯ãä»ããŠããŸãã
OOM ã¯Â Out Of MemoryïŒã¡ã¢ãªäžè¶³ïŒ ã®ç¥ã§ã"ãã®ãã·ã³ã§ã¯èŒããªã" ãšããæå³ã§ãã
LM Studio ã¯ãFull GPU Offload Possibleããšããç·ã®ãããžãåºããŸãããããã¡ãã¯éã«ãç¡çãªãã®ããèŠåããŠããŸãã
å€å®ãæ£ç¢ºãã©ããã¯å¥ãšããŠãèšèšææ³ãšããŠäžæ©èžã¿èŸŒãã§ããå°è±¡ããããŸãã
ã§ã¯ããQwen3.8-27B-GGUFããããããŒãžã§ã³äžèЧãããã§ "recommended" ããŒã¯ã®ãããUD-Q4_K_XLããããŠã³ããŒãããŸãã
ð Unsloth ãæé©åã»é
åžããŠãã GGUFã§ãã³ã³ããã¹ããæšè«æ§èœãèæ
®ãããã¥ãŒãã³ã°ãå
¥ã£ãŠããããŒãžã§ã³ã§ãã

ãUDãã¯ãUnsloth DynamicïŒUnsloth Dynamic éååãé©çšãããã«ãïŒã®ç¥ã§ãã
ã¬ã€ã€ãŒããšã«æé©ãªéååæ¹åŒãéžæããåçéååãæ¡çšããŠãããéèŠãªéšåã¯é«ã粟床ãç¶æãã圱é¿ã®å°ããéšåã¯ãã匷ãå§çž®ããããã«èª¿æŽããã GGUFã«ãªããŸãã

ã¢ãã«ãèªã¿èŸŒããšã次ã®ã¡ãã»ãŒãžãåºãŸããã

å
容ã¯ãmmproj ãGPUã§ã¯ãªãã·ã¹ãã ã¡ã¢ãªã«çœ®ãããç»åå
¥åã¯äœ¿ãããç»ååŠçã¯é
ããªãããšãããã®ã§ãã
ãã®é
眮㯠Redditã®èšå®äŸã§èŠã mmproj-offload = false ãšåãããšãããŠããŸãããã¡ãã¯çµéšã®ãããŠãŒã¶ãŒãæåã§éžãã èšå®ã§ããã
ããããã¡ãã§ã¯æ¢å®ã®æåã«ãªã£ãŠããŸãã
ãªã CPUã«çœ®ãããã®ãã¯åãããŸããã
24GBã«åãŸããªããšå€æããã®ãããããšãæ¢å®ã®æ¹éãªã®ãâŠ
ðMCPãã¹ãçµæ
Unsloth ã® MCPãã¬ãŒã¹ãã«ãã«ãååãšåæ§ã®ã³ãã³ããå ¥åããŸãã
Display name:
filesystemURL or command:
npx -y @modelcontextprotocol/server-filesystem D:\mcp-teståãšåãååã»åãã³ãã³ãã»åãã¹ã³ãŒãã§ãã
ð詳ããã¯åã®èšäºãã芧ãã ããã

å·Šäžã®ãTest connectionããæŒããŠãã¹ããéããã確èªãããã¹ãéãããšã確èªããããã§ããAdd Serversããã¯ãªãã¯ããŸãã

察象ãšãããã©ã«ããååãšåãã«ããŠããŸããããæ¡ä»¶Aãã ãããããš.txt ãäºéã«ä»ã圢ã§å€æãå¿ èŠãªã¿ã€ãã«ã«ããŠããŸãã

çµæãšããŠã
ãæ¡ä»¶A_ã¡ã¢.txt.txtããèŠã€ããŠãç§»åæã«ä¿®æ£ãããšææ¡ããŠããŸããã倿ãå²ããããªå
容ãã¡ãããšæŸã£ãŠä¿®æ£ããŠããŸãã


âïžLM Studioãšã®æåã®éã
ããã§ã®ãããšãã®æåã«ã€ããŠãæ¹ããŠæ³šç®ããŸãã
å
ã»ã©ã® LM Studio ã§ã®ãã¹ãã§ã¯ãåæã«ãã£ãŠããããåäœã§ããã
ã§ãããããã§ã¯éããŸããã
ãã£ãã確èªããŠããŸãã
ãã®å 容ã§é²ããŠããããã§ããïŒãŸããå¥ã®æŽçæ¹æ³ããåžæã§ããããèãããã ããã

ãã€ã¯ããã®æåã¯ã¢ãã«ã®ä»æ§ã§ã¯ãããŸããã
ãã£ããç»é¢å·Šã®ãApprove for meïŒç§ã®ããã«æ¿èªããïŒãèšå®ãå¹ããŠããããã§ãã

ãã®æ©èœããã¢ãã«ã®å¿çèšèšã«ãŸã§åãã§ãããã®ã«ãªããŸãã
èšå®ã¯ãã¯ãªãã¯ããé
ç®ãéžæããããšã§ãèªç±ã«å€ããããšãå¯èœã§ãã

ð åè. éžæå 容ã«ã€ããŠ
ã»Ask for approvalïŒ
ããã¡ã€ã«ç·šéã»ã€ã³ã¿ãŒãããå©çšã®åã«å¿
ã確èª
ã»Approve for meïŒæ¢å®ïŒïŒ
ãå®è¡ããããèªèšŒæ
å ±ã¢ã¯ã»ã¹ã»æš©éææ Œã»ç Žå£çã³ãã³ãã®åã«ã¯ç¢ºèª
ã»Run automaticallyïŒ
ããµã³ãããã¯ã¹å
ãªã確èªãªãã§å®è¡
ã»Full accessïŒ
ãç¡å¶éã確èªãªããã³ãŒããµã³ãããã¯ã¹ãç¡å¹
æäžéšã®ãFull accessïŒç¢ºèªãªãïŒããéžã¹ã°ãäžåºŠã確èªãªãäžæ°ã«äœæ¥ãè¡ããŸãã
â»å®å
šæ§ã®èгç¹ãããFull access èšå®ã¯é¿ããŠãããæ¹ãè¯ãã§ããã確èªããªãåãããªãé«éã«äœæ¥ãå®äºããŸãããã詊ãã§ãã£ãŠã¿ãŠããçµãã£ããæ»ãã®ãå¿ããªãããã«ã
ã€ãŸããLM Studio ãšã¯æš©éèšèšã«å¯Ÿããææ³ãæ£å察ãªä»æ§ã«ãªããŸãã
LM Studioã¯ãäœãæäœãããã³ã«ãæ¬åœã«å®è¡ããŸããïŒããšç¢ºèªããŠããŸãããUnslothã§ã¯èªåã§ãã©ããŸã§èªåã§æ¿èªãããããæ±ºããããšãã§ããŸãã
ããããLM Studioã§ã¯å
ã»ã©ã®ãã¹ãã®ããã«ã確èªãªãã§éãã±ãŒã¹ãããããããã®å¢çãã©ãã§æ·ããŠããã®ãããã¡ãã§ã¯åãããŸããã
ããã«å¯ŸããŠãUnsloth ã¯èªåã§å¢çãæ·ãããšãã§ãããšããããšã«ãªããŸãã
ã»LM Studioã¯ãå¢çããã·ã¹ãã ã®æ¢å®ã®ã«ãŒã«ããæ±ºããŠããèšèš
ã»Unslothã¯ããã®å¢çãã©ãã«çœ®ãããããŠãŒã¶ãŒèªèº«ã決ãããèšèš
ð¥ããã¥ã¡ã³ãã®èªã¿èŸŒã¿
ã§ã¯æ¬¡ã«ãããã¥ã¡ã³ãã®èªã¿èŸŒã¿ãã¹ãã§ãã
Unsloth ã§ã¯ããŒã¿ã®èªã¿èŸŒã¿ã¯Â RAGæ©èœãšããŠæ±ãããŸãã

ãã¹ãããäžã§æ³šæããŠããã¹ãããšã 2ç¹ã ããäŒãããŸãã
â ããã¥ã¡ã³ãã¿ã€ãã«è¡šèš
è±èªã¿ã€ãã«ã¯æ£ãã衚瀺ãããŸããã
æ¥æ¬èªã¿ã€ãã«ã§ã¯ã_.pdfããšãªããŸãã

èšå®ã®ãããŒã¿ããããã¢ããããŒãããããŒã¿ãåé€ãããã¿ç®±ããããŸãïŒäžèšç»åïŒããã¡ã€ã«åã衚瀺ãããªããªããæ°ãå¢ããå Žåã«ãã©ããæ¶ãã°ããã®ãå€å¥ã§ããªããªãããšã«ãªããŸãã
ãã®ããã
Unslothã«æãããã¡ã€ã«ã¯ãäºåã«è±æ°ååã«ãªããŒã ãã
ãšèŠããŠãããšè¯ãããšæããŸãã
â¡ãã£ããå è¡ã®ãããšã
ãŸãããã£ãããå è¡ããŠããããã¥ã¡ã³ããèªã¿èŸŒãŸããæé ããããŸãããã£ãããªãã§æ·»ä»ãå ã«è¡ãããšãããšå€±æããã®ã§ããæ³šæãã ããã

Couldn't attach documents â Could not start a chat to attach them to.
ïŒçŽèš³ïŒ
ããã¥ã¡ã³ããæ·»ä»ã§ããŸããã§ãã â æ·»ä»å ã®ãã£ãããéå§ã§ããŸããã§ããã
Couldn't start a chat for these documents
ïŒçŽèš³ïŒ
ãããã®ããã¥ã¡ã³ãçšã®ãã£ãããéå§ã§ããŸããã§ããã
ã€ãŸãããã£ãããäœæããäžã§è¡ãå¿ èŠãããããšããäºã§ãã
âïž--mmproj-offload ãšã³ã³ããã¹ãåæžã®çµã¿åããã§èª¿æŽ
ãã¹ãçµæããäŒããããšã
ååãšåããå³é¢PDFããèªãŸããŸããããäœåºŠãã¹ããç¹°ãè¿ããŠãã25% ãããããé²ãŸãªããªããŸãã

ç»é¢ã«ã¯ãCPUã§åãããŠããŸãããšåºãŠããŸããã
ãã ãçŽãæ¹ãæžããŠãã£ãã®ã¯Â ãã°ã®äžã§ããã
Auto: running the vision projector on the CPU (--no-mmproj-offload); it does not fit in VRAM alongside the model at 4096 context. Image encoding gets slower.
ïŒçŽèš³ïŒ
Auto: ããžã§ã³ãããžã§ã¯ã¿ãŒãCPUã§å®è¡ããŸãïŒ--no-mmproj-offloadïŒãã³ã³ããã¹ãé·4096ã®å Žåãã¢ãã«ãšåæã«VRAMã«åãŸããªãããã§ãããã®çµæãç»åãšã³ã³ãŒãã¯äœéåããŸãã
ã¢ãã«ãåœåããŠã³ããŒãããéã«ãæ³šææžãã衚瀺ãããŸããããCPUã§åŠçããå Žåã®å€§å¹
ãªé床äœäžããããã»ã©å€§ããªåœ±é¿ãåãŒããšã¯èªèã§ããŠããŸããã§ããã
çŸåšãmmproj ã¯CPUäžã§åäœããŠããŸãã
ãã®ãããç»ååŠçã䌎ãåŠçã¯äœéã«ãªãããšãæ¡å
ãããŠããŸãã
ãã¯ã¿ãŒåœ¢åŒã®ããŒãžã§ããã°ããã¹ãæœåºã®ã¿ã§åŠçã§ããŸããããã®ããŒãžã¯ç»ååŠçãå¿
èŠãªããããã®æ©æµãåããããŸããã
ãã®ãããªç»ååŠçã§ã¯ãCPUãããVRAMïŒGPUïŒãé©åã§ãã
ð¡ è£è¶³. mmproj ãšã¯
multimodal projectorïŒãã«ãã¢ãŒãã«æåœ±åšïŒã®ç¥ã
ããžã§ã³å¯Ÿå¿ã¢ãã«ã¯ãå®ã¯ 2ã€ã®éšåã§ã§ããŠããŸãã
æ¬äœïŒLLMïŒ â ããã¹ããæ±ããQwen3.8-27B ã®å Žå 16.4GB
mmproj â ç»åãæ¬äœãçè§£ã§ãã圢ã«å€æããã0.9GB
ç»åã¯ãã®ãŸãŸã§ã¯èšèªã¢ãã«ã«å
¥ããŸãããmmproj ãç»åããã¯ãã«ã«å€æããããã¹ãã®ããŒã¯ã³ãšåã圢åŒã«ãæåœ±ãããã
ã ãã projector ã§ãã
GGUF圢åŒã§ã¯å¥ãã¡ã€ã«ã§é åžãããŸãã
Unsloth ã§ mmproj-F16 ã¯ãHugging Faceã®ãã£ãã·ã¥ã«ãããŸãã
C:\Users\<ãŠãŒã¶ãŒå>\.cache\huggingface\hub\snapshots

ããã§çŽæ¥ã¢ãã«ã確èªããŠãåãããŸãããmmproj ã¯F16ã§ãBF16 ã§ã¯ãããŸããã
ã€ãŸããæ¬äœã4bitã«èœãšããŠããmmproj ã¯é«ç²ŸåºŠã®ãŸãŸã§ãã
èªåå€å®ã§ã·ã¹ãã RAM ã«çœ®ãããäºïŒ--no-mmproj-offloadïŒãåå ã§ããããšã¯æçœã§ããã®ããç»åãšã³ã³ãŒãã 50ã100åé ããªãã玢åŒãæ¢ãŸã£ãâŠ
LM Studioã§ã¯èª¿æŽãèŠã€ãããŸããã§ããããUnsloth ã§ã¯ãªããšãçµè·¯ã®èª¿æŽãã§ãããã§ãã
ããã§ã
ã³ã³ããã¹ããåã£ãŠ VRAM ã空ãã
mmproj ã GPU ã« offloadããŸãã
ãã ã ãNew downloads onlyã ãšæèšãããŠãããããæ¢åã®ã¢ãã«ã¯ç§»åããŸãããããŸå
¥ã£ãŠãã18GB㯠Cã«æ®ããŸãã
ïŒ
å ·äœçãªèª¿æŽå 容ã¯ä»¥äžã§ãã
1. Context Length ãäžããïŒ37888 â 8192ã16384çšåºŠïŒ
2. Extra Arguments ã« --mmproj-offload ã远å ã


ããã§ãç»åãšã³ã³ãŒããGPUã«èŒããã©ããâŠ
ã¡ãªã¿ã«ãoffload ã¯Â llama.cpp ã®çšèªã§ããGPUã«èŒããããšããæå³ã§ãã
--mmproj-offloadïŒmmproj ã GPU ã«èŒãã
--no-mmproj-offloadïŒGPU ã«èŒãããã·ã¹ãã RAM ã§åãã
èšå®ãæ¹ã㊠ããŒãããŠããŸãç»é¢ã«ãvision projector on the CPUãã®éç¥ãåºãããããã㯠ãšã©ãŒã«ãªããâŠ
å·Šäžã®ãã§ãã¯ããã¯ã¹ãRemember for this modelãã«ãã§ãã¯ããŠãLoad modelãã§ä¿åãçŽããŸãã
ãªããšããããŸã VRAMã«èŒããŸããã

ã³ã³ããã¹ãã¯ã8.2k ãšãããŸãã
37888ããäžããçµæã衚瀺ãããŠããŸãã

KVãã£ãã·ã¥ã 2.9GB â 0.6GBçšåºŠïŒã³ã³ããã¹ãã1/4以äžïŒã«ãªããmmproj ãçŽ0.9GBå¢ããããšããèšç®ããã
å·®ãåŒãGPUã¯Â 18ã19GBå°ã«ãªãã¯ãã§ãã
å°çšGPUã¡ã¢ãªã§ç¢ºèªãããšã19.9GBã
äºæ³ãã18ã19GBå°ã«è¿ãå€ã§ãã

ããã§ãmmproj ã GPUã«èŒã£ããšèããŠè¯ãããã§ãã
ãã£ããã«è©ŠããŠã¿ããšãããéåºŠã¯æ¡éãã
46ç§ã§å®äºããŸããã
çŽ50ã100åïŒ
ããŒããã°ã«ã¯ mmprojã®è¡ãå¢ããŠããŸãã
GGUF size: 16.4 GB, mmproj: 0.9 GB, est. KV cache: 1.1 GB, context: 8192
Appending user extra args: ['--rope-scaling','yarn','--yarn-orig-ctx','32768','--mmproj-offload']åå㯠mmproj: 0.9 GB ã®é
ç®èªäœããããŸããã§ããïŒVRAMåå®ã«å
¥ã£ãŠããªãã£ãïŒã
åŒæ°ãå¹ããŠãmmproj ã GPUåŽã®èšç®ã«å«ãŸããŠããŸãã
16.4 + 0.9 + 1.1 = 18.4GB
ã¿ã¹ã¯ãããŒãžã£ãŒã® 19.9GBãšãæŽåããŸãã
ð¬èªã¿èŸŒãã PDFã®å 容ã§ãã£ãããã
Unslothã§ã¯ãã®ãããªèª¿æŽãå¯èœã§ãããããQwen3.8ã§ãé©åãªåŠççµè·¯ãæ§æããæåã§éãããšãã§ããŸããã
Qwen3.8 ã¯ãçšéã«å¿ããŠçްããªèª¿æŽãå¿ èŠã«ãªãå Žé¢ãå€ããã³ãã¥ããã£ã§ãããŸããŸãªãããšããèŠãããŸãã
éçºè ã¬ãã«ãäžå¿ã§ããããšããããããŒã«ã«LLMã®æ±çšæ§ã¯ãŸã çºå±éäžãšããå°è±¡ã§ããããšãŒãžã§ã³ãç³»AIã¢ãã«ãæå ã®ç°å¢ã§èªç±ã«åãããçŸåšã®ç¶æ³ã«ã¯ã倧ããªå¯èœæ§ãšæåŸ ãæããŸãã
è£è¶³ãšããŠãäŒãããŠãããšã
èªã¿èŸŒãã ããã¥ã¡ã³ãã®å
容ã«ã€ããŠãã£ããããéã¯ãContext Length ã 16384 ãŸã§åŒãäžããç¶æ
ã§åäœã確èªããŸããã
èšç®äž:
ã»16.4ïŒã¢ãã«ïŒ+ 0.9ïŒmmprojïŒ+ 2.2ïŒKV 16384ïŒ= 19.5 GB
ã»VRAM Budget 21.2 GB
ãããŸã§åèå€ã§ã¯ãããŸãããmmproj ãçŽ 0.9 GB ã®VRAMã䜿çšããç¶æ ã§ãã©ããŸã§ Context Length ã䌞ã°ãããã«ã€ããŠã¯ãæçµçã«ã¯å®æž¬ã§ç¢ºèªãããããããŸããã
ãã®èª¿æŽã®ç®å®ã«ã€ããŠã以äžã«èšèŒããŸãã®ã§ãåèã«ããŠäžããã
ð§®äºç®ã®ç«ãŠæ¹ïŒKVãã£ãã·ã¥ïŒ
ã³ã³ããã¹ãã®äºç®ãèããéã¯ããŸã KVãã£ãã·ã¥ãèŠç©ããã®èµ·ç¹ã«ããŸãã
ãKVãã£ãã·ã¥ïŒKey-Value CacheïŒããšã¯ãAttentionã®èšç®ã§çæããã KïŒKeyïŒãšVïŒValueïŒ ãä¿åããåŸç¶ã®çæã¹ãããã§åå©çšããä»çµã¿ã§ãã
åå±€ã®Attentionã§ã¯ãå ¥åãã KãšÂ Vãå°åœ±ããŠçæããããããå±€ããšã«ã¡ã¢ãªãžèç©ããŸããéå»ã®ããŒã¯ã³ã«å¯Ÿå¿ãã Kãš Vã¯åèšç®ããå¿ èŠããªãããããã£ãã·ã¥ããŠããããšã§ã以éã®çæã¹ããããããã®ãŸãŸåç §ã§ããŸãã
ã€ãŸããã³ã³ããã¹ãé·ãå¢ããã»ã©ãKVãã£ãã·ã¥ã«å¿
èŠãªVRAMãå¢ããŠãããŸãã
ãã®ãããã³ã³ããã¹ããå¿
èŠä»¥äžã«å€§ãããããšãKVãã£ãã·ã¥ãVRAMãå§è¿«ããã¢ãã«æ¬äœãVRAMã«åããããªããªãããšããããŸãã
ããã§ãKVãã£ãã·ã¥ã®äºç®ãç«ãŠãŠããã®ã§ãã
äžè¬çã«ãKVãã£ãã·ã¥ã®ãµã€ãºã¯æ¬¡ã®èŠçŽ ã«ãã£ãŠæ±ºãŸããŸãã
ã³ã³ããã¹ãé· â é·ããªãã»ã©æ¯äŸããŠå¢å
ã¢ãã«ã®å±€æ°ã»Attentionãããæ§æ â ã¢ãã«åºæ
KV Cacheã®ããŒã¿åïŒDtypeïŒ â f16ãåºæ¬ãq8_0ã«ãããšãããããååçšåºŠãŸã§åæžå¯èœ
KVãã£ãã·ã¥ã¯ã³ã³ããã¹ãé·ã«å¯ŸããŠã»ãŒæ¯äŸããããããŸãã¯å®æž¬å€ãã1ããŒã¯ã³ãããã®å¢å éãæ±ããŸãã
ä»åã® Qwen3.8-27Bã§ã¯ã
37,888ããŒã¯ã³ â 2.9 GB
8,192ããŒã¯ã³ â 1.1 GB
ãšãã宿ž¬å€ãåŸãããŸããã
ãã®å·®åããã
2.9 â 1.1 = 1.8 GB
ãšãªããŸãã
ã³ã³ããã¹ãé·ã®å·®Â ã¯ã
37,888 â 8,192 = 29,696ããŒã¯ã³
ãªã®ã§ãçŽ29.7kããŒã¯ã³å¢å ããããšã§ãKVãã£ãã·ã¥ãçŽ1.8GBå¢ããŠããŸãã
ããããæŠç®ãããšã
29.696 kïŒKVãã£ãã·ã¥ïŒ ÷ 1.8 GBïŒã³ã³ããã¹ãïŒ â 0.0606 GB / k
ã€ãŸãã
çŽ0.061 GB / 1kããŒã¯ã³
çŽ61 MB / 1kããŒã¯ã³
ãšãªããŸãã
Qwen3.8-27B / KV Cache f16 ã§ã¯ã
1kããŒã¯ã³ãããçŽ0.061 GBïŒçŽ61MBïŒã®å¢å ãšãäºæ³ãç«ãŠãããšãã§ããŸãã
ãã ããåºå®ã®ãªãŒããŒãããã 0.4ã0.5GB ã»ã©ããããã§ãã®ã§ãããã¯ãããŸã§ç®å®ãšããŠèŠãŸãã
ð§®äºç®ã®ç«ãŠæ¹ïŒã³ã³ããã¹ãé·ïŒ
VRAM ããèŠãã³ã³ããã¹ãé·ã®äºç®ã¯ã次ã®åŒã§æŠç®ã§ããŸãã
KV â 0.5 + ã³ã³ããã¹ãé·(k) à 0.064 GB
äŸ
VRAM BudgetïŒ21.2 GBïŒ
â ã¢ãã«æ¬äœïŒ16.4 GBïŒ
â mmprojïŒ0.9 GBïŒ
= 3.9 GB ãKVã«äœ¿ããäžé3.9 GB ãäžã®åŒã«ä»£å
¥ãããšã
0.5 + n à 0.064 = 3.9 â n â 53k
ã€ãŸããçè«äžã¯çŽ53,000ããŒã¯ã³ãŸã§æ ŒçŽã§ããèšç®ã«ãªããŸãã
ãã ãããã®å€ã«ã¯Â ç»åãšã³ã³ãŒãæã«å¿
èŠãšãªãäžæãããã¡ã¯å«ãŸããŠãããããã®åã® VRAMæ¶è²»ã¯å¥éèæ
®ããå¿
èŠããããŸãã
ãŸãããã®åŒã¯ãããŸã§æŠç®ã§ãããå®éã®äœ¿çšå¯èœãªã³ã³ããã¹ãé·ã¯ã¢ãã«ãå®è¡æ¡ä»¶ã«ãã£ãŠå€åããŸãã
ãããããããŸã§ç®å®ã§ãã
æã確å®ãªã®ã¯ããã°ã確èªããããšã§ãã
èšå®ã倿ŽããŠã¢ãã«ããªããŒããããšãUnslothãèªåã§èšç®ãã est. KV cache ã衚瀺ãããŸãã
ãããŸã§ã®èšç®åŒã¯ããããŸã§ãã ããããã®ãããããšããç®å®ãä»ããããã®ãã®ã§ããæçµçãªå€æã¯ãå®éã®ãã°ã«è¡šç€ºãããå€ãåºæºã«ããŠãã ããã
ð€ãªããMuse Glimmerã¯åããŠQwen3.8ã¯åããªãã£ãã®ã
ãããŸã§æ€èšŒããŠããŠãã²ãšã€æ°ã«ãªãããšããããŸãã
ãªããMuse Glimmerã¯LM Studioã§ãé£ãªãæ±ããã®ã«ãQwen3.8-27Bã§ã¯ã³ã³ããã¹ããæº¢ããŠããŸã£ãã®ãã
æåŸã«ãã®éãããŸãšããŸãã
â åã30Bã§ããå¿ èŠãªã¡ã¢ãªã¯åãã§ã¯ãªã
ã©ã¡ãã 30BååŸã®ã¢ãã«ã§ãã
VRAM 24GBãšããåãç°å¢ã§è©ŠããŠãã以äžãæåã¯åçŽã«ãã¢ãã«ãµã€ãºã®éããã ãšæã£ãŠããŸãããããããå®éã«Unslothã§çްããæ°å€ã調æŽããŠãããšãã©ããããã ãã§ã¯èª¬æã§ããŸããã
çµè«ãããããšã
ãã©ã¡ãŒã¿æ°ãè¿ãïŒå¿
èŠ VRAMãè¿ããã§ã¯ãªã
ãšããããšã§ãã
â äžçªå€§ããã®ã¯ãã¢ãã«ã®ãæ§é ãã®éã
Muse Glimmer ã¯çŽ29.6Bã§ãããããªãç¹æ®ãªæ§é ã«ãªã£ãŠããŸãã
ç¹ã«å€§ããã®ã GQAïŒGrouped-Query AttentionïŒã§ãQ/KVãããã 32 / 2ãã€ãŸã16åã® Queryãããã§1ã€ã® KVããããå ±æããŸãã
Meta ã®è³æã§ãããã®æ§é ã«ãã£ãŠ KVãã£ãã·ã¥ã倧ããåæžã§ãããšèª¬æãããŠããŸããããã«ãAttentionãå šå±€ã Full Attentionã§ã¯ãªãã3å±€ã®Sliding Window + 1å±€ã®Full Attention ãšããæ§æã§ãã
ããã«å¯ŸããŠãQwen3.8-27B ã§ã¯ãä»å宿ž¬ããããã«ãã³ã³ããã¹ãã䌞ã°ããš KVãã£ãã·ã¥ãããªã倧ããå¢ããŸãã
Muse Glimmer
ã30Bãªã®ã«KVãã£ãã·ã¥ãæããããæ§é
Qwen3.8-27B
ã27Bã§ãã³ã³ããã¹ãã䌞ã°ããšKVãVRAMãå§è¿«ãã
â ç»ååŠçã®çµè·¯ã®éã
Muse Glimmerã¯æåããããŒã«ã«ã®ãšãŒãžã§ã³ãçšéãæèããŠèšèšãããã¢ãã«ã§ãå°çšã®Perception Encoderãæã£ãŠããŸãã
Metaèªèº«ããconsumer hardwareã§åãããããšãèšèšç®æšã®äžã€ãšããŠæããŠããŸãã
äžæ¹ãä»åã® Qwen3.8ã§ã¯ãåçŽã«ãã¢ãã«æ¬äœãVRAMã«å ¥ãã°OKãã§ã¯ãªãã
ã¢ãã«æ¬äœ
ïŒ KVãã£ãã·ã¥
ïŒ mmproj / ç»ååŠç
ïŒ æšè«ã©ã³ã¿ã€ã ã®ã¯ãŒã¯ã¹ããŒã¹
ïŒ ç»åãšã³ã³ãŒãæã®äžæãããã¡
ãŸã§èããªããšãããŸããã
ã ãããRTX 4090ã®Â 24GBãšããç©çç㪠VRAM容éã¯åãã§ããå®éã«ã¢ãã«ãžæž¡ãããèªç±ãªVRAMãã¯å šç¶éã ããã§ãã
ããã§éèŠã«ãªãã®ãããã©ã¡ãŒã¿æ°ã§ã¯ãªããã¢ãã«ã®æ§é ãã®ãã®ã§ãã
ããã¯åçŽã«ã
ãLM StudioããUnslothã®ã»ããåªããŠããã
ãšãã話ã§ã¯ãããŸããã
ãããã
ã¢ãã«ã«ãã£ãŠãå¿ èŠãšããããåŠççµè·¯ããéãã
ãšããããšã§ãã
ð¯GPUæ§èœãããã¡ã¢ãªæŽ»çšããž
ãã®ãããªããšãããã³ãã¥ããã£ã§ã¯ããŸããŸãªå·¥å€«ãç¶ããŠããŸãã
ããã§æè¿ç®ç«ã€ã®ãããGPUã ãã䞻圹ã«ããªããæ¹åãžã®æé©åã§ãã
ã€ãŸããããã«ããã¯ãåçŽãªæŒç®æ§èœããããã¢ãã«ãã©ã®ã¡ã¢ãªéå±€ã«ã©ãåãããããžç§»ãã€ã€ãã ãšããããšã§ãã
ç¹ã« MoEïŒMixture of ExpertsïŒã¢ãã«ã§ã¯ãæšè«æã«å®éã«äœ¿çšããã®ã¯äžéšã®ãšãã¹ããŒãã ãã§ãããã¢ãã«å šäœãä¿æããããã®ã¡ã¢ãªã¯å¿ èŠã«ãªããŸããããã§ãllama.cpp ã§ã¯éã¢ã¯ãã£ããªãšãã¹ããŒããã·ã¹ãã RAMãžãªãããŒããã --n-cpu-moe ã®ãããªä»çµã¿ãçšæãããŠããŸãã
ããã«ããã³ãœã«åäœã§ VRAMãš RAMãžã®é 眮ã調æŽãã --override-tensor ãªã©ããããããã¹ãŠãVRAMã«èŒãããã®ã§ã¯ãªããå©çšå¯èœãªã¡ã¢ãªå šäœãã©ãçµã¿åãããããšããçºæ³ãåºãã£ãŠããŸãã
ã³ãã¥ããã£ã§ã¯ã極端ãªéååããã³ãœã«é 眮ã®å·¥å€«ã«ãã£ãŠããããŸã§24GBçŽã®VRAMãå¿ èŠãšèããããŠãã35Bã¯ã©ã¹ã®ã¢ãã«ããããå°ãªãVRAMã§åããäºäŸãŸã§ç»å ŽããŠããŸãã

極端ãªåçéåå â 1-bitçŽãŸã§å§çž®ããªãããå®çšçãªç²ŸåºŠãç¶æããæ¹å
ãã³ãœã«é 眮ã®å¶åŸ¡ â ã©ã®å±€ãVRAMã«çœ®ããã©ããã·ã¹ãã RAMãžãªãããŒããããã調æŽ
ã©ã³ã¿ã€ã åŽã®å¯Ÿå¿ â æ°ããã¢ãŒããã¯ãã£ãç»å ŽãããšããŸã llama.cpp åŽã§å¯Ÿå¿ãé²ã¿ããã®åŸ LM Studioã Ollamaãªã©ã«ãå®è£ ãåºãã£ãŠãã
ããããŠèŠããšãçŸåšã®ããŒã«ã«LLMã®æé©åã¯ãåçŽãªããã€ãšã³ãGPUåãã®é«éåãã ãã§ã¯ãªãã
ããããæå
ã«ããVRAMãã·ã¹ãã RAMããããŠãããã®éã®ã¡ã¢ãªåž¯åãå«ãããã¡ã¢ãªéå±€å
šäœã«ãã¢ãã«ãã©ãè©°ã蟌ããããšããæè¡ã«ãªãã€ã€ãããŸãã
ã€ãŸããã巚倧åç«¶äºããããå¹çåç«¶äºããžãšå°ããã€ç§»ã£ãŠãããšæ³šç®ããŸãã
ðããŒã«ã«LLMãä¿ããã·ã³ã®é²å
ã¯ã©ãŠãã®AIãµãŒãã¹ã䜿ã£ãŠãããšãããããéãã¯ã»ãšãã©æèããããšããããŸããã
ã¢ãã«ãéžæããŠãå ¥åããã°ãããã§åããŸãã
ãããããŒã«ã«LLM ã§ã¯äºæ ãéããŸãã
ã©ã®ã¢ãã«ã䜿ãã®ãã
ã©ã®éååãéžã¶ã®ãã
ã©ãã ãã®ã³ã³ããã¹ãã確ä¿ããã®ãã
KVãã£ãã·ã¥ãã©ã®Dtypeã§æã€ã®ãã
ç»ååŠçãã©ãã§è¡ãã®ãã
VRAMãšã·ã¹ãã RAMãã©ã䜿ãåããã®ãã
ãããŠãã©ã®ã©ã³ã¿ã€ã ã䜿ãã®ãã
ããããæ¡ä»¶ã«ãã£ãŠãåã GPUã§ãçµæãå€ãããŸãã
ãããŠããã®ãåŠççµè·¯ãã©ãæ§æãããããšããèãæ¹ã¯ãçŸåšã®ã¡ã¢ãªäŸ¡æ Œã®é«éš°ãšãç¡é¢ä¿ã§ã¯ãªãããã«æããŸãã
ãããŸã§ã® PCãã¯ãŒã¯ã¹ããŒã·ã§ã³ã§ã¯ããã髿§èœãª GPUãæèŒããããšãããã®ãŸãŸæ§èœåäžã«ã€ãªãããšããèãæ¹ãæ¯èŒçåããããããã®ã§ããã
ãããããŒã«ã«LLMã§ã¯ãå°ãæ§å³ãå€ãããŸãã
髿§èœãª GPUãè€æ°æèŒããŠããŠããã¢ãã«ã®æ§é ãå¿ èŠãªã¡ã¢ãªå®¹éã«ãã£ãŠã¯ããã®æ§èœãååã«æŽ»ãããªãã±ãŒã¹ããããŸãã
äžæ¹ã§ãGPUã ãã«äŸåããã倧容éã®ãŠããã¡ã€ãã¡ã¢ãªãæèŒãããã·ã³ã«æ³šç®ãéãŸãã®ããããããèæ¯ãããããã§ãããã
ã€ãŸããããããã®AIåããã·ã³ã§ã¯ã
ãã©ãã ãéãGPUãæã£ãŠãããã
ã ãã§ã¯ãªãã
ãã¢ãã«ãã©ã®ã¡ã¢ãªã«ãã©ã®ããã«é 眮ã§ãããã
ãšããèŠç¹ããããéèŠã«ãªã£ãŠããã®ãããããŸããã
ããã¯ãåãªã GPUæ§èœç«¶äºãšã¯å°ãéãé²åã§ãã
VRAMãã·ã¹ãã RAMããŠããã¡ã€ãã¡ã¢ãªããããŠããããã€ãªãã¡ã¢ãªåž¯åã
éããããªãœãŒã¹ãã©ã®ããã«çµã¿åãããã¢ãã«ãã©ãåããã®ãã
ããŒã«ã«LLMã®æ®åã¯ããœãããŠã§ã¢åŽã®æé©åã ãã§ãªãããAIãåãããã·ã³ãã®ãã®ã®èšèšææ³ãã«ãå€åãä¿ããŠããããã«æããŸãã
- #AI
- #çæAI
- #AI掻çš
- #åç»
- #AIäºæ³
- #AIãšãŒãžã§ã³ã
- #ã»ãã¥ãªãã£
- #aiåç»
- #3DCG
- #ããŒã«ã«LLM
- #ããŒã¯ã³
- #ã¡ã¢ãª
- #gpu
- #ããŒã«ã«AI
- #AIã»ãã¥ãªãã£
- #AIæè¡
- #AIã¢ãã«
- #æšè«
- #ã³ã³ããã¹ã
- #Astra
- #ããŒãã¹
- #GPT6Astra
- #ã³ã³ããã¹ããšã³ãžãã¢ãªã³ã°
- #LMStudio
- #VRAM
- #Qwen3
- #ãªããžããª
- #rtx
- #LocalLLM
- #RTX5090
- #AIã¢ãã«æ¯èŒ
- #unsloth
- #KVãã£ãã·ã¥
- #ãŠããã¡ã€ãã¡ã¢ãª
- #MuseGlimmer
- #ã©ã³ã¿ã€ã
- #ããŒã«ã«AIæ§ç¯
- #NVFP4
- #glimmer
- #mmproj
- #AIã©ã³ã¿ã€ã
- #reasoningbudget
- #gpuoffload