Video wird geladen...

Video konnte nicht geladen werden

Beim Laden dieses Videos ist ein Problem aufgetreten. Dies könnte an einem vorübergehenden Netzwerkproblem liegen oder das Video ist möglicherweise nicht verfügbar.

I didn't expect DeepSeek v4 PRO (not Flash) to run well on the Mac Studio M3 Ultra with 512GB of RAM. This is 2 bit quantized with the same DwarfStar recipe used for Flash. 433GB GGUF file. 130 t/s prefill, 13 t/s generation. Prefill in the video is low... show more

antirez

75,097 subscribers

171,478 Aufrufe • vor 3 Monaten •via X (Twitter)

Gaming Bildung Wissenschaft & Technologie

Anya Rossi• Live Now

Private livecam show

0 Kommentare

Keine Kommentare verfügbar

Kommentare vom Original-Post werden hier angezeigt

Ähnliche Videos

Crazy prefill speed of DwarfStar running DeepSeek v4 PRO in the DGX Station (private branch for now) using techniques that leverage the peculiar RAM + VRAM mixed setup of the Station. 3000 t/s but even more can be reached.

Crazy prefill speed of DwarfStar running DeepSeek v4 PRO in the DGX Station (private branch for now) using techniques that leverage the peculiar RAM + VRAM mixed setup of the Station. 3000 t/s but even more can be reached.

antirez

62,001 Aufrufe • vor 17 Tagen

DeepSeek v4 PRO running in DwarfStar on a DGX Station. Can go faster, already peaks 50 t/s without DSpark: dynamic VRAM/RAM allocation of experts based on past statistics, 4k t/s prefill on big chunks. VRAM/RAM layer-major streaming. Mixed RAM/VRAM inference requires care.

DeepSeek v4 PRO running in DwarfStar on a DGX Station. Can go faster, already peaks 50 t/s without DSpark: dynamic VRAM/RAM allocation of experts based on past statistics, 4k t/s prefill on big chunks. VRAM/RAM layer-major streaming. Mixed RAM/VRAM inference requires care.

antirez

26,246 Aufrufe • vor 18 Tagen

DGX Station: this is just the start of what it is possible to do (and why I claim DwarfStar could be *the* Station inference engine). DeepSeek v4 PRO Q2 with routed experts split among VRAM / RAM with kernels optimized for this peculiar setup. 45 t/s but can go faster.

DGX Station: this is just the start of what it is possible to do (and why I claim DwarfStar could be the Station inference engine). DeepSeek v4 PRO Q2 with routed experts split among VRAM / RAM with kernels optimized for this peculiar setup. 45 t/s but can go faster.

antirez

41,545 Aufrufe • vor 19 Tagen

Still didn't attempt any quality testing, but let's say that GPT 5.6 Sol was able almost unassisted to write this Laguna S2.1 inference implementation following the other two models in DwarfStar, and it is very fast, 50 t/s generation, very fast prefill as well.

Still didn't attempt any quality testing, but let's say that GPT 5.6 Sol was able almost unassisted to write this Laguna S2.1 inference implementation following the other two models in DwarfStar, and it is very fast, 50 t/s generation, very fast prefill as well.

antirez

43,547 Aufrufe • vor 1 Monat

DwarfStar DeepSeek v4 Flash 0731 pre-M5 kernel keeps improving 💪🏻 - prefill +11% - decode +10% - single request no DSpark/MTP. - 100% same logics between baseline and now. Today is QA like crazy. I hope to file PR by EOD antirez Here are results on the M3 Ultra and a video now vs baseline using ds4-eval.

DwarfStar DeepSeek v4 Flash 0731 pre-M5 kernel keeps improving 💪🏻 - prefill +11% - decode +10% - single request no DSpark/MTP. - 100% same logics between baseline and now. Today is QA like crazy. I hope to file PR by EOD antirez Here are results on the M3 Ultra and a video now vs baseline using ds4-eval.

Ivan Fioravanti ᯅ

17,121 Aufrufe • vor 28 Tagen

I'm now used to this speed 🤣 Running DeepSeek-V4-Flash-IQ2XXS on my single 3090 and 64GB RAM tg is fine, its the prefill that actually hurts.

I'm now used to this speed 🤣 Running DeepSeek-V4-Flash-IQ2XXS on my single 3090 and 64GB RAM tg is fine, its the prefill that actually hurts.

AJ

20,999 Aufrufe • vor 2 Monaten

GLM 4.7 Flash is supported in mlx-lm 0.30.3 (h/t Ivan Fioravanti ᯅ) The 4-bit runs fast (43 tok/s generation, ~800 tok/s prefill) on a base M5 32GB laptop.

GLM 4.7 Flash is supported in mlx-lm 0.30.3 (h/t Ivan Fioravanti ᯅ) The 4-bit runs fast (43 tok/s generation, ~800 tok/s prefill) on a base M5 32GB laptop.

Awni Hannun

141,194 Aufrufe • vor 7 Monaten

First kinda working implementation of GLM 5.2 in DwarfStar. Will take some time to be good enough, but it is a promising start. 433 GB GGUF file. M3 Ultra 512GB.

First kinda working implementation of GLM 5.2 in DwarfStar. Will take some time to be good enough, but it is a promising start. 433 GB GGUF file. M3 Ultra 512GB.

antirez

69,565 Aufrufe • vor 2 Monaten

1-bit GLM-5.2 GGUF vs. Claude 4.8 Opus vs. GPT-5.5 We gave 3 models the same prompt and compared one-shot outputs. The 1-bit GLM-5.2 GGUF ran locally on a Mac Studio M3 Ultra with 256GB RAM at ~21.6 tok/s. Which output do you like best? GGUF:

1-bit GLM-5.2 GGUF vs. Claude 4.8 Opus vs. GPT-5.5 We gave 3 models the same prompt and compared one-shot outputs. The 1-bit GLM-5.2 GGUF ran locally on a Mac Studio M3 Ultra with 256GB RAM at ~21.6 tok/s. Which output do you like best? GGUF:

Unsloth AI

1,543,147 Aufrufe • vor 2 Monaten

GLM 5.2 flash Q2 (generated by myself with similar recipe of DS4 Flash of DwarfStar GGUF) running on an M5 Max with 128GB. The Q4 weights also work on tensor parallel inference in two M5 Max systems. Soon on GitHub. Still testing + CUDA + ROCm ...

GLM 5.2 flash Q2 (generated by myself with similar recipe of DS4 Flash of DwarfStar GGUF) running on an M5 Max with 128GB. The Q4 weights also work on tensor parallel inference in two M5 Max systems. Soon on GitHub. Still testing + CUDA + ROCm ...

antirez

28,103 Aufrufe • vor 8 Tagen

Deepseek V4 Flash 0731 (Q2) - 12 tokens/sec - Single RTX 4090 - 650+ tokens/sec prefill - 250k context - no kv cache quantization! DeepSeek just dropped the official V4 Flash 0731 two days ago with a massive agent capabilities upgrade. The official benchmarks are literally crushing their own V4-Pro-Preview on agentic tasks like Terminal Bench 2.1 and DeepSWE. Unsloth AI said they couldn't wait to bring it to local devices, and they delivered. If you thought my 118B Poolside Laguna S 2.1 MoE run last week on a single GPU was wild, hold onto your hardware. I just successfully ran Unsloth’s brand new 91GB DeepSeek-V4-Flash-0731 (UD-IQ2_M) GGUF entirely locally. And I pushed it to a mind-bending 250,000 context window. The VRAM ceiling is an illusion if you know how to optimize llama.cpp. Here are the benchmarks and the cheat codes to run a local frontier class model yourself. For the hardware and setup, I used a single NVIDIA RTX 4090 (24GB VRAM) hooked up via a PCIe 4 bus, running Ubuntu 22.04 LTS and CUDA 13.0. You don't need a massive enterprise server for this, if you have more than 80 GB of standard DDR4 RAM and a 24GB card like an RTX 3090 or 4090, you can run this exact stack yourself. All benchmarks were run using a massive 28k token prompt to truly stress test the prefill limits. no kv cache quantization THE BENCHMARKS (Scaling Context): # 80k Context (Baseline: -b 2048 -ub 2048): Prefill: 465.43 t/s | Decode: 13.00 t/s | VRAM: 22.87 GB # 80k Context (Optimized: -b 4096 -ub 4096): Prefill: 643.15 t/s | Decode: 12.20 t/s | VRAM: 23.00 GB (Notice how doubling the batch flags spiked my prefill throughput by nearly 200 t/s with almost zero VRAM penalty) # 180k Context (-b 4096 -ub 4096): Prefill: 629.18 t/s | Decode: 11.92 t/s | VRAM: 23.40 GB # 250k Context MAXIMUM (-b 4096 -ub 4096): Prefill: 619.02 t/s | Decode: 11.54 t/s | VRAM: 23.40 GB # THE SECRET SAUCE (Why this works): Unsloth’s UD-IQ2_M quant is ~91GB across 3 files. Since I only have 24GB of VRAM, the PCIe 4 bus and system RAM have to do the heavy lifting. The magic bullet is the --no-mmap flag. By completely bypassing OS disk paging, I forced llama.cpp to load the massive model weights directly into the system RAM upfront. Combined with Flash Attention (-fa on) and exactly 12 CPU threads (--threads 12), I maintained an incredibly stable 11.5+ tokens/sec decode speed even at a quarter million token context. # THE EXACT COMMAND: ./build/bin/llama-server -m /workspace/models/DeepSeek-V4-Flash-0731-UD-IQ2_M-00001-of-00003.gguf -c 250000 -fa on --port 8080 --threads 12 -b 4096 -ub 4096 --no-mmap -v Local conversational and agentic coding AI is fully here. You don’t need an API or an H100 cluster. Qwen 3.8 27b drops next week making the 24GB VRAM tier even more worthwhile. What does your current local AI rig look like, and what's the craziest model you've managed to squeeze into it? Official huggingface GGUF links from Unsloth and performance graphs are dropped in the replies below!

Alok

46,022 Aufrufe • vor 1 Monat

DeepSeek-V4-Flash vs Gemini-3.6-Flash both given the exact same prompt to build 3D pirate ship Cost: > deepseek-v4-flash: $0.14 / $0.28 > gemini-3.6-flash: $1.50 / $7.50 though deepseek is 20x cheaper, gemini was surprising faster with the output Prompt used: "build a 3D model of pirate ship in Three js with minimal dashboard buttons in light theme background, with wide view of the pirate ship"

DeepSeek-V4-Flash vs Gemini-3.6-Flash both given the exact same prompt to build 3D pirate ship Cost: > deepseek-v4-flash: $0.14 / $0.28 > gemini-3.6-flash: $1.50 / $7.50 though deepseek is 20x cheaper, gemini was surprising faster with the output Prompt used: "build a 3D model of pirate ship in Three js with minimal dashboard buttons in light theme background, with wide view of the pirate ship"

Tim Jayas

20,127 Aufrufe • vor 1 Monat

The VRAM barrier is officially dead. I just ran Qwen 3.8 Flash Next (MoE) 125B A6B with a 250,000 context window on a single 24GB RTX 4090. 21 tokens/sec decode. 364 t/s prefill. no mtp. no dflash. no kv cache quantization! We are running datacenter models on consumer hardware. Tested on Ubuntu 22 | CUDA 13.0 | PCIe 4.0 x16 | 110 GB DDR4 System RAM with a continuous 28k prompt across all runs. ### The Benchmarks & Scaling # 1. Hybrid Offload (-ncmoe 40 @ 80k Context) Offloaded 40 expert layers to the GPU, pushing VRAM to the ceiling. ./build/bin/llama-server -m Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf -c 80000 --port 8080 -v --fit off -b 4096 -ub 4096 -ncmoe 40 Prefill: 383.85 t/s | Decode: 22.52 t/s Footprint: 23.85 GB VRAM | 97 GB RAM # 2. Full CPU MoE Offload (-cmoe @ 80k Context) Pinned all 512 expert layers to DDR4 RAM (-cmoe), keeping attention on the 4090. llama.cpp flags: (Same as above, replace -ncmoe 40 with -cmoe) Prefill: 355.72 t/s | Decode: 20.84 t/s Footprint: 11.66 GB VRAM (12GB+ VRAM freed up!) | 110 GB RAM # 3. The 180,000 Context Run Prefill: 357.75 t/s | Decode: 20.98 t/s | VRAM: 15.6 GB | RAM: 110 GB # 4. The 250,000 Context Absolute Ceiling ./build/bin/llama-server -m Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf -c 250000 --port 8080 -v --fit off -b 4096 -ub 4096 -cmoe Prefill: 364.29 t/s | Decode: 20.97 t/s Footprint: 18.3 GB VRAM (Still ~5.7 GB of VRAM headroom!) | 110 GB RAM ### Key Insights: -b 4096 -ub 4096: doubles the prompt ingestion from ~150 to 364+ t/s. -cmoe Free Lunch: Shifting expert layers to DDR4 RAM slashes VRAM from 24GB to 11.6GB with virtually zero decode penalty (22.5 -> 20.9 t/s), enabling the 250k context ceiling. Qwen 3.8 Flash-Next (UD-Q4_K_XL) is a massive 111.4 GB model split across 4 shards. To run this architecture, you must build from the experimental PR branch (#27742) by Daniel Han: git clone && cd llama.cpp git fetch origin pull/27742/head:qwen-next && git checkout qwen-next cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native -DBUILD_SHARED_LIBS=OFF cmake --build build --config Release -j $(nproc) --target llama-server A single 4090 paired with 100 GB of cheap DDR4 RAM will comfortably serve production grade 125B inference. While Qwen 3.8 27B (dense) still holds the crown for single 3090/4090 rigs, Flash Next proves 125B hybrid models are officially viable on consumer hardware. Hugging Face GGUF link and complete performance telemetry graphs are dropped in the replies below. GLM 5.3 Flash VS Qwen 3.8 Flash Next, which one takes the open weights crown this week?

The VRAM barrier is officially dead. I just ran Qwen 3.8 Flash Next (MoE) 125B A6B with a 250,000 context window on a single 24GB RTX 4090. 21 tokens/sec decode. 364 t/s prefill. no mtp. no dflash. no kv cache quantization! We are running datacenter models on consumer hardware. Tested on Ubuntu 22 | CUDA 13.0 | PCIe 4.0 x16 | 110 GB DDR4 System RAM with a continuous 28k prompt across all runs. ### The Benchmarks & Scaling # 1. Hybrid Offload (-ncmoe 40 @ 80k Context) Offloaded 40 expert layers to the GPU, pushing VRAM to the ceiling. ./build/bin/llama-server -m Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf -c 80000 --port 8080 -v --fit off -b 4096 -ub 4096 -ncmoe 40 Prefill: 383.85 t/s | Decode: 22.52 t/s Footprint: 23.85 GB VRAM | 97 GB RAM # 2. Full CPU MoE Offload (-cmoe @ 80k Context) Pinned all 512 expert layers to DDR4 RAM (-cmoe), keeping attention on the 4090. llama.cpp flags: (Same as above, replace -ncmoe 40 with -cmoe) Prefill: 355.72 t/s | Decode: 20.84 t/s Footprint: 11.66 GB VRAM (12GB+ VRAM freed up!) | 110 GB RAM # 3. The 180,000 Context Run Prefill: 357.75 t/s | Decode: 20.98 t/s | VRAM: 15.6 GB | RAM: 110 GB # 4. The 250,000 Context Absolute Ceiling ./build/bin/llama-server -m Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf -c 250000 --port 8080 -v --fit off -b 4096 -ub 4096 -cmoe Prefill: 364.29 t/s | Decode: 20.97 t/s Footprint: 18.3 GB VRAM (Still ~5.7 GB of VRAM headroom!) | 110 GB RAM ### Key Insights: -b 4096 -ub 4096: doubles the prompt ingestion from ~150 to 364+ t/s. -cmoe Free Lunch: Shifting expert layers to DDR4 RAM slashes VRAM from 24GB to 11.6GB with virtually zero decode penalty (22.5 -> 20.9 t/s), enabling the 250k context ceiling. Qwen 3.8 Flash-Next (UD-Q4_K_XL) is a massive 111.4 GB model split across 4 shards. To run this architecture, you must build from the experimental PR branch (#27742) by Daniel Han: git clone && cd llama.cpp git fetch origin pull/27742/head:qwen-next && git checkout qwen-next cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native -DBUILD_SHARED_LIBS=OFF cmake --build build --config Release -j $(nproc) --target llama-server A single 4090 paired with 100 GB of cheap DDR4 RAM will comfortably serve production grade 125B inference. While Qwen 3.8 27B (dense) still holds the crown for single 3090/4090 rigs, Flash Next proves 125B hybrid models are officially viable on consumer hardware. Hugging Face GGUF link and complete performance telemetry graphs are dropped in the replies below. GLM 5.3 Flash VS Qwen 3.8 Flash Next, which one takes the open weights crown this week?

Alok

1,000,881 Aufrufe • vor 9 Tagen

For the DGX Spark owners. This is what you get with DS4 in your hardware. I want to post this to show how with fast prefill and not very fast generation, the system remains absolutely fine to use.

For the DGX Spark owners. This is what you get with DS4 in your hardware. I want to post this to show how with fast prefill and not very fast generation, the system remains absolutely fine to use.

antirez

40,085 Aufrufe • vor 3 Monaten

Initial DeepSeek v4 and v4 Flash support added to anemll-flash-llama.cpp! 🚀 M5 MAX 128GB can run full DeepSeek-V4 with 1.6T params in original FP8/FP4 weights from SSD without requantization! One-click scripts to convert HF safetensors to dense + MoE sidecar (no GGUF needed) Inference and server examples Prefill and decoding benchmarks Branch and docs here: This is WIP and I plan to add cleanup and more testing this week. Feedback welcome!

Initial DeepSeek v4 and v4 Flash support added to anemll-flash-llama.cpp! 🚀 M5 MAX 128GB can run full DeepSeek-V4 with 1.6T params in original FP8/FP4 weights from SSD without requantization! One-click scripts to convert HF safetensors to dense + MoE sidecar (no GGUF needed) Inference and server examples Prefill and decoding benchmarks Branch and docs here: This is WIP and I plan to add cleanup and more testing this week. Feedback welcome!

Anemll

32,250 Aufrufe • vor 4 Monaten

DS4 running on DGX Spark (GB10 / CUDA), private branch for now. 12 tokens/sec, the memory bandwidth is limited in this system, at 270GB/sec. But prefill is ways more alighed to M3 Max at ~200 t/s. I'll release when more mature, but it is almost sure that it will get merged.

DS4 running on DGX Spark (GB10 / CUDA), private branch for now. 12 tokens/sec, the memory bandwidth is limited in this system, at 270GB/sec. But prefill is ways more alighed to M3 Max at ~200 t/s. I'll release when more mature, but it is almost sure that it will get merged.

antirez

84,127 Aufrufe • vor 3 Monaten

Final results for DeepSeek-V4-Flash 284B 3107 on 4x3090 (2-bit quant, my llama.cpp fork) after all optimizations. This morning: 433 tok/s prefill. Right now: 1906 tok/s. At 98K depth, not empty context. Yes I was not mistaken. Its x4.3 Decode untouched, 37-40 tok/s. We did not change a single model weight. One day of removing three walls between the CPU and the cards. Chronology in the thread, raw server log in the video.

Final results for DeepSeek-V4-Flash 284B 3107 on 4x3090 (2-bit quant, my llama.cpp fork) after all optimizations. This morning: 433 tok/s prefill. Right now: 1906 tok/s. At 98K depth, not empty context. Yes I was not mistaken. Its x4.3 Decode untouched, 37-40 tok/s. We did not change a single model weight. One day of removing three walls between the CPU and the cards. Chronology in the thread, raw server log in the video.

Alexey Fateev

19,063 Aufrufe • vor 1 Monat

Run Updated Gemma 4 26B A4B QAT (MoE) with Vision at 25 tokens/sec and massive 120k context window on a single RTX 4060 (8 GB VRAM + 16 GB RAM Only!!) Yesterday I pushed Gemma 4 26B A4B QAT to 250k context on a single RTX 4060 using nothing but Q8 KV cache and optimized -b and -ub flags for higher prefill throughput. Today I stacked Multi Token Prediction (MTP) self speculative decoding AND the vision projector (mmproj) on top of that same card, same batch size optimization, same $250 GPU and pushed it until it broke, then found the fix. All text only runs consist of a 28k prompt. vision runs consist of 28k text prompt + an image. # 1. MTP alone. near free decode speed, no catch MTP draft assistant is a separate small model (MTP heads are backed into the main model itself for the qwen 3.5+ models but its a separate small model for gemma 4 series), 240 MB gguf 80k ctx: Prefill 510 t/s | Decode 29.5 t/s 120k ctx: Prefill 433 t/s | Decode 29 t/s 180k ctx: Prefill 240 t/s | Decode 24.9 t/s 250k ctx: Prefill 63 t/s | Decode 13 t/s llama.cpp flags: m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --spec-type draft-mtp -md mtp-gemma-4-26B-A4B-it.gguf-c 180000 -b 1024 -ub 1024 --spec-draft-n-max 6 --spec-draft-p-min 0.7 -ctk q8_0 -ctv q8_0 # 2. Add vision on top. the tax you actually pay the vision projector gguf is about 1.1 GBs 80k ctx: Prefill 360 t/s | Decode 25.4 t/s 120k ctx: Prefill 230 t/s | Decode 23.8 t/s 180k ctx (Q8 KV): Prefill 75 t/s | Decode 12.5 t/s - cliff flags: -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --spec-type draft-mtp -md mtp-gemma-4-26B-A4B-it.gguf -c 80000 --port 8080 -b 1024 -ub 1024 --spec-draft-n-max 6 --spec-draft-p-min 0.7 -ctk q8_0 -ctv q8_0 --mmproj mmproj-F16.gguf # 3. The fix if you want to run vision over 120k context: swap Q8 KV for Q4 KV past 120k Stack MTP + vision + Q8 KV past 120k context and you hit a wall. draft model overhead plus KV pressure tanks everything. Drop to Q4 KV and the wall disappears: 180k ctx (Q4 KV): Prefill 220 t/s | Decode 25.5 t/s -ctk q4_0 -ctv q4_0 --mmproj mmproj-F16.gguf (rest same as above) Bottom line: MTP gives you a near free +20-30% decode boost up to 120k context. Past that, it's fighting your VRAM, not helping and if vision is loaded too, Q4 KV isn't optional past 120k, it's mandatory. 30% boost is model and card specific, MTP boosted decode 2x for gemma 4 31b on a single rtx 4090. Same 8GB card. Same $250 GPU. Multimodal, speculative decoding, 180k usable context, zero upgrades. You gotta try this if you have a single NVIDIA RTX 3050, 3060, 3070, 4050, 4060, 5050 or 5060. You can try it with a 6 GB VRAM card as well but you will have to lower the context window. Hugging Face links to the updated Unsloth's QAT quants and performance graph are in the replies below. Which models are you running on your 6/8/12GB cards with MTP?

Run Updated Gemma 4 26B A4B QAT (MoE) with Vision at 25 tokens/sec and massive 120k context window on a single RTX 4060 (8 GB VRAM + 16 GB RAM Only!!) Yesterday I pushed Gemma 4 26B A4B QAT to 250k context on a single RTX 4060 using nothing but Q8 KV cache and optimized -b and -ub flags for higher prefill throughput. Today I stacked Multi Token Prediction (MTP) self speculative decoding AND the vision projector (mmproj) on top of that same card, same batch size optimization, same $250 GPU and pushed it until it broke, then found the fix. All text only runs consist of a 28k prompt. vision runs consist of 28k text prompt + an image. # 1. MTP alone. near free decode speed, no catch MTP draft assistant is a separate small model (MTP heads are backed into the main model itself for the qwen 3.5+ models but its a separate small model for gemma 4 series), 240 MB gguf 80k ctx: Prefill 510 t/s | Decode 29.5 t/s 120k ctx: Prefill 433 t/s | Decode 29 t/s 180k ctx: Prefill 240 t/s | Decode 24.9 t/s 250k ctx: Prefill 63 t/s | Decode 13 t/s llama.cpp flags: m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --spec-type draft-mtp -md mtp-gemma-4-26B-A4B-it.gguf-c 180000 -b 1024 -ub 1024 --spec-draft-n-max 6 --spec-draft-p-min 0.7 -ctk q8_0 -ctv q8_0 # 2. Add vision on top. the tax you actually pay the vision projector gguf is about 1.1 GBs 80k ctx: Prefill 360 t/s | Decode 25.4 t/s 120k ctx: Prefill 230 t/s | Decode 23.8 t/s 180k ctx (Q8 KV): Prefill 75 t/s | Decode 12.5 t/s - cliff flags: -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --spec-type draft-mtp -md mtp-gemma-4-26B-A4B-it.gguf -c 80000 --port 8080 -b 1024 -ub 1024 --spec-draft-n-max 6 --spec-draft-p-min 0.7 -ctk q8_0 -ctv q8_0 --mmproj mmproj-F16.gguf # 3. The fix if you want to run vision over 120k context: swap Q8 KV for Q4 KV past 120k Stack MTP + vision + Q8 KV past 120k context and you hit a wall. draft model overhead plus KV pressure tanks everything. Drop to Q4 KV and the wall disappears: 180k ctx (Q4 KV): Prefill 220 t/s | Decode 25.5 t/s -ctk q4_0 -ctv q4_0 --mmproj mmproj-F16.gguf (rest same as above) Bottom line: MTP gives you a near free +20-30% decode boost up to 120k context. Past that, it's fighting your VRAM, not helping and if vision is loaded too, Q4 KV isn't optional past 120k, it's mandatory. 30% boost is model and card specific, MTP boosted decode 2x for gemma 4 31b on a single rtx 4090. Same 8GB card. Same $250 GPU. Multimodal, speculative decoding, 180k usable context, zero upgrades. You gotta try this if you have a single NVIDIA RTX 3050, 3060, 3070, 4050, 4060, 5050 or 5060. You can try it with a 6 GB VRAM card as well but you will have to lower the context window. Hugging Face links to the updated Unsloth's QAT quants and performance graph are in the replies below. Which models are you running on your 6/8/12GB cards with MTP?

Alok

16,266 Aufrufe • vor 1 Monat

Mistral Vibe CLI running Devstral-small-2-2512 on M3 Ultra 512GB with LM Studio as backend 🔥 First part of the video 1x and the remaining 2x Great speed overall, 27 toks/s as average tracked on LM Studio console. Details to run it below 🧵

Mistral Vibe CLI running Devstral-small-2-2512 on M3 Ultra 512GB with LM Studio as backend 🔥 First part of the video 1x and the remaining 2x Great speed overall, 27 toks/s as average tracked on LM Studio console. Details to run it below 🧵

Ivan Fioravanti ᯅ

49,682 Aufrufe • vor 8 Monaten