I_am_Bishal (@C15C01337) header
I_am_Bishal (@C15C01337) profile picture

Security Research Engineer 💂 CTF Lead w/@hackasec 🕸️ Local LLM • GPU tuning • Chasing tokens/sec 👀 Blackhat MEA 2023/24/25 and Midnight Flag CTF Finalist 🎩

NP
Joined June 2017
🎉 Five years ago today (16 Oct 2020), I played my first-ever CTF solo - unknowingly starting something much bigger. What began as one person’s challenge became @hackasec , Nepal’s #1 cybersecurity CTF team and a global contender. #hackasec #CTF #Cybersecurity #Global #ctftime
4
643
I spent a few hours running Qwen3.8-27B NVFP4 on an RTX PRO 6000 with SGLang 0.5.19 same target weights, three ways to decode: - raw (no draft) - DFlash2 (z-lab draft, 8-token blocks) - DSpark (RadixArk draft) I tried math, coding, prose, and chat, at short context, ~7K, and ~29K. Three runs per setting, medians only and greedy, 512 tokens out each time, temp 0. Raw is basically flat no matter what you ask it. Short prompts, ~58 tok/s ~7K in, ~51 tok/s ~29K in, ~40 tok/s Minimum baseline where Math, code, story, chat roughly the same range. Turn on speculation and coding at short context speed up. Coding, short (~28 tokens in) raw: 58 DFlash2: 171 DSpark: 172 Math, short (~62 tokens in) raw: 58 DFlash2: 137 DSpark: 166 So, ~3× on code if that’s your workload. But math already splits: DSpark ahead of DFlash2 on the short prompt. A single speculative decoding speedup figure would be misleading. Prose and conversational workloads don’t see the same performance gains. Prose, short (~41 tokens in) raw: 58 DFlash2: 108 DSpark: 110 Chat, short (~90 tokens in) raw: 59 DFlash2: 95 DSpark: 90 Still faster than raw but not the 3x as others. Chat has template tokens baked in; creative text doesn’t draft as cleanly as "Answer:" code prompts. Long-context workloads expose the limitations of these performance claims. Prefill becomes more expensive, and speculative decoding acceptance rates decline. Coding, ~29K in raw: 40 DFlash2: 73 DSpark: 68 Math, ~29K in raw: 40 DFlash2: 78 DSpark: 71 Prose, ~29K in raw: 40 DFlash2: 59 DSpark: 57 Chat, ~29K in raw: 40 DFlash2: 60 DSpark: 53 Still ~1.5-1.9× over raw on the stuff that matters to me (code/math at long context). If you stare at SGLang’s accept rate, it clicks. Coding with DFlash2: ~68% short, ~61% at 29K. Prose more like ~30%. Chat can dip to ~18% at 7K on DFlash2. When draft acceptance drops, the system needs more verification rounds, so the actual speedup takes a hit. There’s nothing wrong with the model, it’s just that this type of prompt isn’t really what the draft model is optimized for. DFlash2 vs DSpark? There’s no clear winner across every workload. DSpark has a slight edge in short coding tasks, while DFlash2 performs better on longer math and coding prompts in my tests. It really comes down to what kind of workload you're running. Same GPU, same 27B model. The real difference comes down to the draft model and the type of requests you're serving. Depending on those, speculative decoding can either give you a noticeable speed boost or barely make a difference. Huge thanks to all the teams behind these amazing projects for making it possible to experiment, benchmark, and explore what’s possible. Used Models: - huggingface.co/RadixArk/Qwen… - huggingface.co/z-lab/Qwen3.8… - huggingface.co/RadixArk/Qwen… Recipe: MiaAI-Lab Qwen3.8 RTX PRO 6000 SGLang - github.com/MiaAI-Lab/Qwen3.8… Engine: SGLang + Qwen3.8 cookbook - github.com/sgl-project/sglan… - docs.sglang.io/cookbook/auto… #LLM #SGLang #Qwen #speculativedecoding
4
4
234
10, 100 or 1,000 teammates. $0 seat fees. Introducing OpenDesign Go Team—a low-cost AI subscription for every team, in OpenDesign & via API. $8 with up to $220 in model usage. Shared credits. Usage tracked by teammate. Invite friends to earn Reset Cards together now.
64
47
8
48
27,164
While searching for the smaller model with the top performances came accross Spark-X2.6-4B with 605K+ reported downloads and claims of coding and agent capabilities competitive with larger models. I wanted to see what this compact model could deliver in practice testing raw versus DSpark speed, memory overhead, and inputs up to 256K tokens. How much does a drafter help a full 4B model? I benchmarked Spark-X2.6-4B with and without its official DSpark drafter on one RTX PRO 6000. Both modes had a 1,048,576-token context window allocated. Actual input prompts ranged from 8K to 256K tokens. The biggest gain: 2.86× native decode throughput on math at 256K input, with approximately 1.77 GiB of additional sampled peak GPU memory. Here’s the breakdown. MATH - the strongest gains Draft acceptance stayed between 45.8% and 49.0%. The highest measured median decode rate in the entire test was 318.56 tok/s on math at 16K. Memory: DSpark added 1,808 MiB approximately 1.77 GiB to sampled peak GPU memory above the raw baseline in every task/depth group. That is measured runtime overhead, including additional allocations required by the speculative path. It is not just the drafter’s weight-file size. Memory was sampled device-wide every 100 ms, so shorter transient peaks could be missed. Prefill: DSpark’s prefill throughput was approximately 12–32% lower. For example, at 256K input on math: • Raw prefill: 6,628.5 tok/s • DSpark prefill: 5,777.8 tok/s • Raw decode: 75.94 tok/s • DSpark decode: 217.24 tok/s That distinction matters. A short answer following a huge input can spend most of its time processing the prompt. A 2.86× decode gain should not be presented as a 2.86× improvement in complete-request latency. Setup I used was: • RTX PRO 6000 Blackwell 96 GB Vram • Exact full Spark-X2.6-4B.gguf target in both modes • Raw baseline loaded no drafter • Official DSpark drafter Spark-X2.6-4B-DSpark.gguf with 7 draft tokens • F16 KV caches, Flash Attention and CUDA graphs enabled • Batch 2048 / microbatch 512 • One inference slot, thinking disabled • Greedy target decoding • Two prompt cases per task, three rounds • 512 generated tokens per measured request • No prompt-cache hits • 288 measured requests in total Rates are medians across six measurements per task/depth/mode. Acceptance is pooled accepted/proposed draft tokens. Limitation: only 57 of 144 paired greedy outputs matched exactly. The longer inputs were task-specific synthetic backgrounds, including generated conversation transcripts for chat. This was not a real-world long-conversation or retrieval-quality evaluation. My takeaway: DSpark’s strongest decode gains here were on math and code. Chat and prose benefited less, and long-input prefill introduced a separate cost. The useful question is more specific than “How many tokens/sec?” It’s: “For which workload, at what input depth, with how much extra memory and what happens to the whole request?” Credit to @SparkLLM team for building a capable 4B model and releasing the DSpark drafter alongside it giving the community more to run, test, and build on. Used: - huggingface.co/XHToken/Spark… - huggingface.co/XHToken/Spark…
2
7
308
I_am_Bishal retweeted
We're giving away a Hogwarts Legacy custom wrapped NVIDIA GeForce RTX 5090 to celebrate its official release with DLSS 5 🪄 To enter: 🟢 Share this post 🟢 Comment #RTXON Terms and Conditions: nvidia.com/en-us/geforce/con…
84,650
72,106
7,708
67,499
6,028,763
DLSS 5 の対応を記念して、『ホグワーツ・レガシー』をテーマにしたカスタム GeForce RTX 5090 をプレゼント🎁 応募方法: 🟢 このアカウントをフォロー 🟢 投稿をシェア 🟢 #RTXON とコメント 応募規約: nvda.ws/4AGaERt
18,641
18,956
2,099
16,266
1,611,512
Over 1.5 million downloads last month DavidAU’s Qwen3.8-27B Turbo Fable Cold Fusion caught my attention, so I benchmarked its Q8_0 version with a new DSpine drafter on an RTX PRO 6000 Blackwell. Here’s the setup and what I measured. Hardware: • One RTX PRO 6000 Blackwell, 96 GB Models: • Target: Qwen3.8-27B-TurboFCFusion-735-882-Here-Uncen-NEO-CODER-MAX-Q8_0.gguf • Drafter: Tencent DSpine-b16, converted to GGUF by @ItsmeAjayKV • Engine: DSpine llama.cpp fork, compiled for Blackwell DSpine Q8_0 weights and Q4_K_M where the target stayed Q8. Runtime settings stayed the same: • 7 draft tokens • Batch 512 / microbatch 128 • Q8_0 KV caches for both models • Flash Attention + CUDA graphs • One inference slot, thinking disabled Q8_0-drafter results across input depths: • Short, ~68 tokens: 134.97 tok/s decode • 7K: 151.50 tok/s decode • 29K: 112.02 tok/s decode Q4_K_M-drafter results across input depths: • Short, ~70 tokens: 148.01 tok/s decode • 7K: 157.42 tok/s decode • 29K: 160.51 tok/s decode At 29,000 input tokens: Q8 drafter → Q4 drafter • Median decode: 112.02 → 160.51 tok/s • Draft acceptance: 43.1% → 60.3% • Median prefill: 2,504.8 → 2,502.9 tok/s Prefill barely changed. The improvement showed up in generation, alongside higher draft acceptance. The Q4 drafter also reduced the draft weight file from 2.20 GB to 1.26 GB. A smaller drafter was worth testing. The main model never had to leave Q8. Credit to DavidAU, @TencentHunyuan DSpine team and @ItsmeAjayKV . Still chasing tokens/sec. ⚡ Used: - huggingface.co/aj9o9/Qwen3.8… - huggingface.co/DavidAU/Qwen3…
4
1
6
383
Qwen3.8-Flash-Next, Unsloth UD-Q2_K_XL, running through Strata with a single RTX PRO 6000 Blackwell I aimed for 150 tokens/sec got 300.3 tok/s ⚡ My setup: • 1× RTX PRO 6000 Blackwell, 96 GB VRAM • Strata 0.2.12 / engine 0.1.58 / CUDA 13 • 256K context configured, int8 KV • Single stream, drafting on auto Median prefill / decode, in tok/s: • Short input → 935.8 / 300.3 • 7K input → 7,088.6 / 311.9 • 29K input → 7,561.1 / 271.2 Three runs per depth, 512 generated tokens each. All nine completed with zero cached input tokens. Longest input tested so far is 29K. Next up, pushing deeper into that 256K window. Credit to the @coldniko - Strata and @UnslothAI teams. Anyone running this pack on other hardware? Would love to compare results. Used: - github.com/Niko1221/Strata - huggingface.co/unsloth/Qwen3…
1
3
280
I_am_Bishal retweeted
💥 Dual R9700 just hit ~229 tok/s on Qwen3.8-27B. Engine is Saber (Rust) + DFlash2. Author TTFT ~47ms, prefill 5k+. One card still works: prose ~67, code ~148. Caveat: repo self-test, 210W cap — don't treat this as a floor on your box. via GGZ14 / saber-engine 👇 github.com/GGZ14/saber-engin…
8
4
1
31
2,353
I_am_Bishal retweeted
⚡Meet Qwen3.8-Flash, a multimodal MoE and an early preview of the Qwen4 architecture, now open-weight! The production version Qwen3.8-Flash will be available soon via QwenCloud API at just $ 0.16/1M input tokens and $ 0.47/1M output tokens. 125B parameters + 51B N-gram embeddings, with just 6B activated per token. Unmatched cost-efficiency. What's new: 🥳 - Next architecture: GDN + QSA hybrid attention, Gated Residual, N-gram Embedding & Muon optimizer, serving as a precursor to the architecture used in Qwen4. - Dramatically lower training and inference costs: trained at just 1/9 the cost of Qwen3.7-Plus, while outperforming it across the board with especially strong gains in coding and office tasks. - Strong performance: scoring 58.7 on DeepSWE 1.1, 62.5 on SWE-bench Pro, 73.9 on CoWorkBench, 84.5 on AndroidWorld, and 95.7 on MathVision (with CI). - 262K native context, extensible to 1M with YaRN. We’re also releasing the weights for Qwen3.8-Flash-Next, giving the community an early look at the new architecture we’re exploring for Qwen4.🚀 We can't wait to see what you build with Qwen3.8-Flash!👀👇 - Blog: qwen.ai/blog?id=qwen3.8-flas… - Technical Report: github.com/QwenLM/Qwen3.8-Fl… - Hugging Face: huggingface.co/Qwen/Qwen3.8-… - ModelScope: modelscope.cn/models/Qwen/Qw…
540
1,046
576
8,187
2,126,118
Target: Stacking up to 1 Billion+ daily DeepSeek-v4.1 Flash-1017 tokens for free🔥 Known so far: Terminal Agent: - OpenCode - Cline - Freebuff - Qoder CLI API Key: - beatapi.io - tokenharbor.ai Any others are offering, Save this and Drop them below! 👇💬
1
6
565
For security and privacy, always run this in an isolated environment, such as a VirtualBox virtual machine or a VPS. Also review the platform’s data retention policies and logging practices. More free API Key: - apinex.bond - zenmux.ai/invite/TYJFK0
1
187
Not a deepseek but, it seems it was a great deal by @Zai_org GLM 5.3 Flash is actually pretty good! Been experimenting with it, and I'm impressed so far. The best part? You can try it for FREE on Zcode. - 100M tokens as a welcome gift - 300M tokens for Weekend Builds with just 2 invites That's a lot of free tokens to experiment, build, and test things out. Definitely worth giving it a shot👀 zcode.z.ai/invite/7N9XPQ9D
97
I_am_Bishal retweeted
this blog shows some of the worst bugs claude found, including a zero-click rce on the linux networking stack that roots a server over the network with no creds or interaction, and tens of full rce chains in chrome (recording below is the one in the blog): anthropic.com/research/disco…
9
52
10
504
103,479
I_am_Bishal retweeted
ctf in 2014: run this program in a debugger ctf in 2026: become a threat actor and exploit this 0day with no public poc
Congrats to @killa who solved stealmybitcoin.com hours ago but I screwed up which key went on the server (the flag was a root wallet WIF instead of the branch WIF). Check out their awesome writeup here: github.com/KillaBoi/StealMyB…
16
49
4
1,749
53,055
I_am_Bishal retweeted
HITCON CTF 2026 Be like:
7
1
96
3,971
I_am_Bishal retweeted
Introducing PwnEscape 🦉: a full KVM guest-to-host escape via a Linux kernel networking flaw affecting industry standard hypervisors. CVE-2026-99999 went from untrusted VM code to host-kernel memory reads, controlled writes, and root command execution. More below:
🤖 Made with AI
16
222
46
1,277
394,544
We had an amazing time at #Pwn2Own Ireland 2026 - our 40th Pwn2Own ever! We awarded $1,262,000 for almost 100 0-days. Catch all the highlights at youtu.be/qDTSScQ67Mg?is=CdmJ…
1
9
1
60
9,158
I_am_Bishal retweeted
🏁 Hackअस्त्र CTF 2026 has officially concluded! 🚩 An incredible journey from the Qualifiers to the Finals. 🔥 455 teams 👥 1100+ participants The online Qualifiers narrowed the competition down to the Top 12 Finalist Teams, who advanced to the Hackअस्त्र CTF Finals 2026.
2
1
7
151
2025 BC (before Chatgpt/Claude/Codex/Cursor)🥺
"Dad, what was it like playing CTFs before AI?"
3
263