Загрузка видео...

Не удалось загрузить видео

Возникла проблема при загрузке этого видео. Это может быть связано с временными проблемами сети или видео может быть недоступно.

На главную

DFlash speculative decoding on Apple Silicon Qwen3.5-9B bf16 · M5 Max · greedy exact match ▸ 85 tok/s, 3.3× at 1024 tokens (runtime) ▸ ~70 tok/s, 2.6× in the video (terminal I/O overhead) ▸ 80 tok/s, 3.1× at 2048 tokens (runtime) Currently working on: → Long context (speedup degrades... show more

bstn 👁️

1,460 subscribers

36,942 просмотров • 2 месяцев назад •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

Комментарии: 0

Нет доступных комментариев

Здесь появятся комментарии из оригинального поста

Похожие видео

DFlash v0.1.4 : custom Metal verify kernels for quantized Qwen3 hybrid models, plus significant peak memory reduction at long context. M5 Max 40-core GPU, 64GB, stock mlx_lm baseline: Qwen3.6-35B-A3B-4bit: ► @ 1024 · 138.3 → 300.3 tok/s (2.20x) ► @ 2048 · 135.6 → 246.4 tok/s (1.81x) ► @ 4096 · 134.5 → 208.4 tok/s (1.56x) ► @ 8192 · 133.2 → 177.4 tok/s (1.33x) Qwen3.5-27B-4bit: ► @ 1024 · 33.5 → 79.0 tok/s (2.37x) ► @ 2048 · 33.1 → 70.2 tok/s (2.12x) ► @ 4096 · 31.5 → 55.7 tok/s (1.77x) ► @ 8192 · 33.9 → 45.3 tok/s (1.34x) Working on making this usable for agentic workloads goal is to never drop below baseline at any context depth. LLM decode is memory-bandwidth bound. M5 Max runs at 614 GB/s, that's 1.5x more than M1-M4 Max (400-410 GB/s). Results will vary on lower bandwidth chips.

DFlash v0.1.4 : custom Metal verify kernels for quantized Qwen3 hybrid models, plus significant peak memory reduction at long context. M5 Max 40-core GPU, 64GB, stock mlx_lm baseline: Qwen3.6-35B-A3B-4bit: ► @ 1024 · 138.3 → 300.3 tok/s (2.20x) ► @ 2048 · 135.6 → 246.4 tok/s (1.81x) ► @ 4096 · 134.5 → 208.4 tok/s (1.56x) ► @ 8192 · 133.2 → 177.4 tok/s (1.33x) Qwen3.5-27B-4bit: ► @ 1024 · 33.5 → 79.0 tok/s (2.37x) ► @ 2048 · 33.1 → 70.2 tok/s (2.12x) ► @ 4096 · 31.5 → 55.7 tok/s (1.77x) ► @ 8192 · 33.9 → 45.3 tok/s (1.34x) Working on making this usable for agentic workloads goal is to never drop below baseline at any context depth. LLM decode is memory-bandwidth bound. M5 Max runs at 614 GB/s, that's 1.5x more than M1-M4 Max (400-410 GB/s). Results will vary on lower bandwidth chips.

bstn 👁️

23,120 просмотров • 2 месяцев назад

dflash-mlx v0.1.7 is out. Big adaptive-runtime update, still focused mostly on Qwen3.6 27B 4-bit. @ 2048 tokens, M5 Max, stock mlx_lm baseline: ► 1024: 33.26 → 98.05 tok/s (x2.95) ► 2048: 32.34 → 90.67 tok/s (x2.81) ► 4096: 30.58 → 93.55 tok/s (x3.06) ► 8192: 26.03 → 79.12 tok/s (x3.04) ► 16384: 21.50 → 60.77 tok/s (x2.78) Main change: adaptive verify got a lot smarter. Instead of blindly trying to verify large 16-token blocks all the time, DFlash now watches acceptance + tokens/cycle + real cycle cost. When the draft gets weaker, it drops to smaller 4-token blocks, then probes back up only when the recent cycles make sense. In practice: less wasted verify work, better long-context behavior, and much more useful metrics to understand what is happening. ► retuned adaptive verify for long-context / agentic decode ► richer metrics: tokens/cycle, adaptive block state, CopySpec counters ► /metrics now has real decode avg + logical/real/restored prefill rates ► AIME25 benchmark suite with exact integer scoring ► Qwen thinking default now follows tokenizer/request behavior ► GDN recurrent exactness fixes I also started running AIME25-style long generations. Even around 45k generated tokens, I was still seeing ~40 tok/s on 27B 4-bit. Over the next few days I’ll share more demos: AIME runs, real OpenCode game/project sessions, and full metrics along the way. Still optimizing hard for 27B 4-bit first, while working on custom kernels per Apple GPU generation so more machines can benefit.

dflash-mlx v0.1.7 is out. Big adaptive-runtime update, still focused mostly on Qwen3.6 27B 4-bit. @ 2048 tokens, M5 Max, stock mlx_lm baseline: ► 1024: 33.26 → 98.05 tok/s (x2.95) ► 2048: 32.34 → 90.67 tok/s (x2.81) ► 4096: 30.58 → 93.55 tok/s (x3.06) ► 8192: 26.03 → 79.12 tok/s (x3.04) ► 16384: 21.50 → 60.77 tok/s (x2.78) Main change: adaptive verify got a lot smarter. Instead of blindly trying to verify large 16-token blocks all the time, DFlash now watches acceptance + tokens/cycle + real cycle cost. When the draft gets weaker, it drops to smaller 4-token blocks, then probes back up only when the recent cycles make sense. In practice: less wasted verify work, better long-context behavior, and much more useful metrics to understand what is happening. ► retuned adaptive verify for long-context / agentic decode ► richer metrics: tokens/cycle, adaptive block state, CopySpec counters ► /metrics now has real decode avg + logical/real/restored prefill rates ► AIME25 benchmark suite with exact integer scoring ► Qwen thinking default now follows tokenizer/request behavior ► GDN recurrent exactness fixes I also started running AIME25-style long generations. Even around 45k generated tokens, I was still seeing ~40 tok/s on 27B 4-bit. Over the next few days I’ll share more demos: AIME runs, real OpenCode game/project sessions, and full metrics along the way. Still optimizing hard for 27B 4-bit first, while working on custom kernels per Apple GPU generation so more machines can benefit.

bstn 👁️

16,334 просмотров • 1 месяц назад

Update: Qwen3.5:9b head-to-head (MLX, Apple Silicon optimized) Mac Studio M2 Ultra: 89.74 tok/s Mac Mini M4: 20.82 tok/s MLX basically doubles the speed on both machines.

Update: Qwen3.5:9b head-to-head (MLX, Apple Silicon optimized) Mac Studio M2 Ultra: 89.74 tok/s Mac Mini M4: 20.82 tok/s MLX basically doubles the speed on both machines.

stevibe

134,256 просмотров • 4 месяцев назад

GLM-5.1-478B-NVFP4 Running on: - 4x RTX Pro 6000 - Sglang - 370,000 max tokens (1.75x full context) - p10 27.7 | p90 45.6 tok/s decode (gen) - 1340 tok/s prefill I could get 2x decode if I limit to 64k context (100 tok/s) In this video it operates Figma (:

GLM-5.1-478B-NVFP4 Running on: - 4x RTX Pro 6000 - Sglang - 370,000 max tokens (1.75x full context) - p10 27.7 | p90 45.6 tok/s decode (gen) - 1340 tok/s prefill I could get 2x decode if I limit to 64k context (100 tok/s) In this video it operates Figma (:

0xSero

74,772 просмотров • 2 месяцев назад

I added KV caching and INT8 KV quantization to our transformer inference, improving throughput by 35x. All of this was done from scratch in Rust + CUDA, on top of a homemade ML framework. On a 4-token prompt with 252 generated tokens: - Original: 0.76 tok/s - KV cache fp32: 27.21 tok/s - KV cache int8 (quantized): 27.29 tok/s Try it out yourself here: In practice: - KV caching gave us about a 35x end-to-end speedup - INT8 KV cache kept roughly the same speed as fp32 but cut KV cache memory by 3.78x FP32 cache used 4.5 MB in this run while the INT8 cache used only 1.19 MB This simple change to inference created a huge impact on performance. To learn more about the KV cache and other optimizations like this, check out the blog at

I added KV caching and INT8 KV quantization to our transformer inference, improving throughput by 35x. All of this was done from scratch in Rust + CUDA, on top of a homemade ML framework. On a 4-token prompt with 252 generated tokens: - Original: 0.76 tok/s - KV cache fp32: 27.21 tok/s - KV cache int8 (quantized): 27.29 tok/s Try it out yourself here: In practice: - KV caching gave us about a 35x end-to-end speedup - INT8 KV cache kept roughly the same speed as fp32 but cut KV cache memory by 3.78x FP32 cache used 4.5 MB in this run while the INT8 cache used only 1.19 MB This simple change to inference created a huge impact on performance. To learn more about the KV cache and other optimizations like this, check out the blog at

Reese Chong

52,588 просмотров • 2 месяцев назад

Qwen3.5 runs quite well in mlx-lm. Awesome that we have a frontier-level hybrid model. The context gets longer but the inference speed and memory use barely change. Here's the Q4 generating a space invaders game on an M3 Ultra. Generated 4,120 tokens at 37.6 tok/s.

Qwen3.5 runs quite well in mlx-lm. Awesome that we have a frontier-level hybrid model. The context gets longer but the inference speed and memory use barely change. Here's the Q4 generating a space invaders game on an M3 Ultra. Generated 4,120 tokens at 37.6 tok/s.

Awni Hannun

48,673 просмотров • 4 месяцев назад

DeepSeek-R1-0528-5bit on MLX pushing M3 Ultra 512GB to its limits! 501GB used mem visibile on mactop in the video! Context: 4K tokens Prompt: 190.29 t/s Gen: 11.37 t/s Peak Mem: 487.48 GB! THIS IS APPLE MLX!

DeepSeek-R1-0528-5bit on MLX pushing M3 Ultra 512GB to its limits! 501GB used mem visibile on mactop in the video! Context: 4K tokens Prompt: 190.29 t/s Gen: 11.37 t/s Peak Mem: 487.48 GB! THIS IS APPLE MLX!

Ivan Fioravanti ᯅ

102,949 просмотров • 1 год назад

MiniMax M2.1 in 4-bit cruises on an M3 Ultra with mlx-lm. Generated a space invaders game using 5098 tokens at 47.2 tok/sec:

MiniMax M2.1 in 4-bit cruises on an M3 Ultra with mlx-lm. Generated a space invaders game using 5098 tokens at 47.2 tok/sec:

Awni Hannun

93,826 просмотров • 6 месяцев назад

Researchers found a way to make LLMs 8.5x faster! (without compromising accuracy) Speculative decoding is quite an effective way to address the single-token bottleneck in traditional LLM inference. A small "draft" model first generates the next several tokens, then the large model verifies all of them at once in a single forward pass. If a token at any position is wrong, you keep everything before it and restart from there. This never does worse than normal decoding. But current drafters in Speculative decoding still guess one token at a time. That makes the drafting step itself a bottleneck, capping real-world speedups at 2-3x. DFlash is a new technique that swaps the autoregressive drafter with a lightweight block diffusion model that guesses all tokens in one parallel shot. Drafting cost stays flat no matter how many tokens you speculate. On top of that, the drafter is conditioned on hidden features pulled from multiple layers of the target model and injected into every draft layer, so it makes significantly better guesses than a drafter working from scratch. In the side-by-side demo below, vanilla decoding runs at 48.5 tokens/sec. DFlash hits 415 tokens/sec on the same model, with zero quality loss. It's already integrated with vLLM, SGLang, and Transformers, with draft models on HuggingFace for several models like Qwen3, Qwen3.5, Llama 3.1, Kimi-K2.5, gpt-oss, and many more. I have shared the GitHub repo in the replies! KV caching is another must-know technique to boost LLM inference. I recently wrote an article about it. Read it below. 👉 Over to you: What use case are you working on that can benefit from this new technique?

Researchers found a way to make LLMs 8.5x faster! (without compromising accuracy) Speculative decoding is quite an effective way to address the single-token bottleneck in traditional LLM inference. A small "draft" model first generates the next several tokens, then the large model verifies all of them at once in a single forward pass. If a token at any position is wrong, you keep everything before it and restart from there. This never does worse than normal decoding. But current drafters in Speculative decoding still guess one token at a time. That makes the drafting step itself a bottleneck, capping real-world speedups at 2-3x. DFlash is a new technique that swaps the autoregressive drafter with a lightweight block diffusion model that guesses all tokens in one parallel shot. Drafting cost stays flat no matter how many tokens you speculate. On top of that, the drafter is conditioned on hidden features pulled from multiple layers of the target model and injected into every draft layer, so it makes significantly better guesses than a drafter working from scratch. In the side-by-side demo below, vanilla decoding runs at 48.5 tokens/sec. DFlash hits 415 tokens/sec on the same model, with zero quality loss. It's already integrated with vLLM, SGLang, and Transformers, with draft models on HuggingFace for several models like Qwen3, Qwen3.5, Llama 3.1, Kimi-K2.5, gpt-oss, and many more. I have shared the GitHub repo in the replies! KV caching is another must-know technique to boost LLM inference. I recently wrote an article about it. Read it below. 👉 Over to you: What use case are you working on that can benefit from this new technique?

Avi Chawla

157,390 просмотров • 1 месяц назад

The RTX 3090 is a 5-year-old GPU and it still runs a 27B model at 20 tok/s I tested Qwen3.5:27b across 3 generations of NVIDIA: 5090 → ~60 tok/s 4090 → ~40 tok/s 3090 → ~20 tok/s Perfectly linear scaling. Double the generation, double the speed.

The RTX 3090 is a 5-year-old GPU and it still runs a 27B model at 20 tok/s I tested Qwen3.5:27b across 3 generations of NVIDIA: 5090 → ~60 tok/s 4090 → ~40 tok/s 3090 → ~20 tok/s Perfectly linear scaling. Double the generation, double the speed.

stevibe

142,379 просмотров • 3 месяцев назад

Today's autoregressive models generate one token at a time. Mercury 2 generates tokens in parallel. Over 1,000 tok/sec on standard GPUs, at comparable quality to speed-optimized models. Since launch, the community has been showing what diffusion LLMs can unlock. Thanks to the team at Clyep for the breakdown.

Today's autoregressive models generate one token at a time. Mercury 2 generates tokens in parallel. Over 1,000 tok/sec on standard GPUs, at comparable quality to speed-optimized models. Since launch, the community has been showing what diffusion LLMs can unlock. Thanks to the team at Clyep for the breakdown.

Inception

21,104 просмотров • 1 месяц назад

Yay, llama2.c can now load and inference the Meta released models! :) E.g. here inferencing the smallest 7B model at ~3 tokens/s on 96 OMP threads on a cloud Linux box. Still just CPU, fp32, one single .c file of 500 lines: expecting ~300 tok/s tomorrow :)

Yay, llama2.c can now load and inference the Meta released models! :) E.g. here inferencing the smallest 7B model at ~3 tokens/s on 96 OMP threads on a cloud Linux box. Still just CPU, fp32, one single .c file of 500 lines: expecting ~300 tok/s tomorrow :)

Andrej Karpathy

409,394 просмотров • 2 лет назад

Sequential Monte Carlo speculative decoding from Makora keeps multiple draft tokens alive in parallel instead of rewinding failed matches.

Sequential Monte Carlo speculative decoding from Makora keeps multiple draft tokens alive in parallel instead of rewinding failed matches.

SemiAnalysis

24,216 просмотров • 26 дней назад

Apple Silicon + Gemma 4 fans: this is for you. Pico AI Server now supports continuous batching with MLX-Swift. 43 tok/s on 1 stream. 26 tok/s per stream on 2 concurrent streams. That’s 52 tok/s total. a 21% throughput gain on a six-year-old MacBook Pro M1 Max!

Apple Silicon + Gemma 4 fans: this is for you. Pico AI Server now supports continuous batching with MLX-Swift. 43 tok/s on 1 stream. 26 tok/s per stream on 2 concurrent streams. That’s 52 tok/s total. a 21% throughput gain on a six-year-old MacBook Pro M1 Max!

Ronald Mannak

48,116 просмотров • 2 месяцев назад

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 просмотров • 5 месяцев назад

nvidia's 3B mamba destroyed alibaba's 3B deltanet on the same RTX 3090. only 24 days between releases. same active parameters, same VRAM tier, completely different architectures. nemotron cascade 2: 187 tok/s. flat from 4K to 625K context. zero speed loss. flags: -ngl 99 -np 1. that's it. no context flags, no KV cache tricks. auto-allocates 625K. qwen 3.5 35B-A3B: 112 tok/s. flat from 4K to 262K context. zero speed loss. flags: -ngl 99 -np 1 -c 262144 --cache-type-k q8_0 --cache-type-v q8_0. needed KV cache quantization to fit 262K. both models held a flat line across every context level. both architectures are context-independent. but nvidia's mamba2 is 67% faster at generating tokens on the exact same hardware and needs fewer flags to get there. same node, same GPU, same everything. the only variable is the model. gold medal math olympiad winner running at 187 tokens per second on single RTX 3090 a card from 6 years ago. nvidia cooked.

nvidia's 3B mamba destroyed alibaba's 3B deltanet on the same RTX 3090. only 24 days between releases. same active parameters, same VRAM tier, completely different architectures. nemotron cascade 2: 187 tok/s. flat from 4K to 625K context. zero speed loss. flags: -ngl 99 -np 1. that's it. no context flags, no KV cache tricks. auto-allocates 625K. qwen 3.5 35B-A3B: 112 tok/s. flat from 4K to 262K context. zero speed loss. flags: -ngl 99 -np 1 -c 262144 --cache-type-k q8_0 --cache-type-v q8_0. needed KV cache quantization to fit 262K. both models held a flat line across every context level. both architectures are context-independent. but nvidia's mamba2 is 67% faster at generating tokens on the exact same hardware and needs fewer flags to get there. same node, same GPU, same everything. the only variable is the model. gold medal math olympiad winner running at 187 tokens per second on single RTX 3090 a card from 6 years ago. nvidia cooked.

Sudo su

186,388 просмотров • 3 месяцев назад

MLXs CUDA backend is getting better. It's especially nice if you appreciate fast startup times. But it's also quite fast in general. Here's Qwen3 4B in fp8 running on my DGX Spark. - Processed 18.5k tokens in < 4 seconds - Generates at 32.5 tok/sec with 18.5k context

MLXs CUDA backend is getting better. It's especially nice if you appreciate fast startup times. But it's also quite fast in general. Here's Qwen3 4B in fp8 running on my DGX Spark. - Processed 18.5k tokens in < 4 seconds - Generates at 32.5 tok/sec with 18.5k context

Awni Hannun

16,542 просмотров • 4 месяцев назад

How Fast is Gemma 4 on a MacBook Pro M4? Benchmarking Google's new MoE (26B-A4B) > Model size: 26.1 GiB > Load time: ~4.2s Comparing single request VS > concurrent requests performance > 32k total context, 4 parallel slots single request behavior > TTFT: 5.68s > prompt: 3,701 tokens @ 652 tok/s > decode: 40.08 tok/s sequential (1 request at a time): > avg duration: 20.5s > p99: 22.1s > throughput: 40.11 tok/s > clean finishes: 100% concurrent (4 parallel requests): > aggregate throughput: 47.25 tok/s > total system throughput: 262.27 tok/s > avg duration: 65.1s > p95 latency: 68.8s > req/sec: 0.058 Head-to-Head: Sequential vs Concurrent throughput: > 40.11 tok/s → 47.25 tok/s (+17.8%) > small gain despite 4x parallelism latency per request: > 20.5s → 65.1s (~3.2x slower) > you pay heavily for concurrency system throughput (true utilization): > ~40 tok/s → 262 tok/s (~6.5x total output) > this is where concurrency wins tokens per second (decode ceiling): > ~40 tok/s steady in both modes > hardware-bound, not scheduler-bound TTFT impact: > ~5.7s baseline → buried under queueing in concurrent > “headers waittime” becomes the bottleneck What this actually means? - You don’t get linear scaling from parallel slots - You trade latency for total output - Mac Unified Memory setup is clearly saturating - Bandwidth + Scheduling overhead show up immediately This is exactly why GPUs dominate here Concurrency without killing latency

How Fast is Gemma 4 on a MacBook Pro M4? Benchmarking Google's new MoE (26B-A4B) > Model size: 26.1 GiB > Load time: ~4.2s Comparing single request VS > concurrent requests performance > 32k total context, 4 parallel slots single request behavior > TTFT: 5.68s > prompt: 3,701 tokens @ 652 tok/s > decode: 40.08 tok/s sequential (1 request at a time): > avg duration: 20.5s > p99: 22.1s > throughput: 40.11 tok/s > clean finishes: 100% concurrent (4 parallel requests): > aggregate throughput: 47.25 tok/s > total system throughput: 262.27 tok/s > avg duration: 65.1s > p95 latency: 68.8s > req/sec: 0.058 Head-to-Head: Sequential vs Concurrent throughput: > 40.11 tok/s → 47.25 tok/s (+17.8%) > small gain despite 4x parallelism latency per request: > 20.5s → 65.1s (~3.2x slower) > you pay heavily for concurrency system throughput (true utilization): > ~40 tok/s → 262 tok/s (~6.5x total output) > this is where concurrency wins tokens per second (decode ceiling): > ~40 tok/s steady in both modes > hardware-bound, not scheduler-bound TTFT impact: > ~5.7s baseline → buried under queueing in concurrent > “headers waittime” becomes the bottleneck What this actually means? - You don’t get linear scaling from parallel slots - You trade latency for total output - Mac Unified Memory setup is clearly saturating - Bandwidth + Scheduling overhead show up immediately This is exactly why GPUs dominate here Concurrency without killing latency

Ahmad

88,866 просмотров • 3 месяцев назад

Local AI can be so good, but you’d need about 12k USD to get it. Then it’s not so great Here’s a Q4 of Qwen3.5-262B-REAP Weights 131GB KVCache 50GB 256,000 context 350 tokens/s prefill warmup 4,000 tokens/s prefill cache 36 tokens/s generation Vision enabled REAP is good

Local AI can be so good, but you’d need about 12k USD to get it. Then it’s not so great Here’s a Q4 of Qwen3.5-262B-REAP Weights 131GB KVCache 50GB 256,000 context 350 tokens/s prefill warmup 4,000 tokens/s prefill cache 36 tokens/s generation Vision enabled REAP is good

0xSero

87,959 просмотров • 3 месяцев назад

llama.cpp with MTP support makes local models fast enough to use as daily drivers 🚀 Qwen3.6-27B dense generation (on A10G): From 25 tok/s → 45 tok/s (+78%). Two flags on llama-server: --spec-type draft-mtp --spec-draft-n-max 2

llama.cpp with MTP support makes local models fast enough to use as daily drivers 🚀 Qwen3.6-27B dense generation (on A10G): From 25 tok/s → 45 tok/s (+78%). Two flags on llama-server: --spec-type draft-mtp --spec-draft-n-max 2

Victor M

171,310 просмотров • 1 месяц назад