Видео с ютуба Skildcore

The SSM Layer Nobody Made Fast (Until Now): ReplaySSM Explained

The SSM Layer Nobody Made Fast (Until Now): ReplaySSM Explained
Bolt-On vs Native Drafters: Medusa, EAGLE, and MiniMax DSpark Architecture

Bolt-On vs Native Drafters: Medusa, EAGLE, and MiniMax DSpark Architecture
700 tok/s on a 2.8 Trillion Model: The 92-Layer TPU Megakernel

700 tok/s on a 2.8 Trillion Model: The 92-Layer TPU Megakernel
How vLLM's NIXL Push Mode Cuts Disaggregated TTFT by 3× (RDMA Write Deep Dive)

How vLLM's NIXL Push Mode Cuts Disaggregated TTFT by 3× (RDMA Write Deep Dive)
How SGLang Fixed the Hybrid Attention Memory Cliff (Attention + Mamba Serving)

How SGLang Fixed the Hybrid Attention Memory Cliff (Attention + Mamba Serving)
The Green Dashboard Outage: Why Disaggregated LLMs Stall Over TCP

The Green Dashboard Outage: Why Disaggregated LLMs Stall Over TCP
Prefill vs Decode: Why LLM Inference Is So Slow (GPU Bottlenecks Explained)

Prefill vs Decode: Why LLM Inference Is So Slow (GPU Bottlenecks Explained)
The 72-Hour Parity War: DeepSeek DeepInfer vs. vLLM vs. SGLang

The 72-Hour Parity War: DeepSeek DeepInfer vs. vLLM vs. SGLang
NVIDIA and AMD Just Shipped Custom Kernels for the Same Engine (On the Same Day)

NVIDIA and AMD Just Shipped Custom Kernels for the Same Engine (On the Same Day)
Why LLMs Read 140 GB per Token — Speculative Decoding & Tree Attention

Why LLMs Read 140 GB per Token — Speculative Decoding & Tree Attention
Why Remote KV Caching Slowed Down Our LLM (The 2.53 GB/s Break-Even Rule)

Why Remote KV Caching Slowed Down Our LLM (The 2.53 GB/s Break-Even Rule)
How vLLM's Megakernel Cuts MoE Decode Latency by 50% (Kimi K3 Breakdown)

How vLLM's Megakernel Cuts MoE Decode Latency by 50% (Kimi K3 Breakdown)
Why a 1-Trillion Parameter Model Dies at the Optimizer (Muon vs AdamW)

Why a 1-Trillion Parameter Model Dies at the Optimizer (Muon vs AdamW)
Inside DeepSeek's 155 Trillion Token Engine: What They Didn't Open Source

Inside DeepSeek's 155 Trillion Token Engine: What They Didn't Open Source
How LLMs Generate Tokens Faster: Speculative Decoding + Tree Attention

How LLMs Generate Tokens Faster: Speculative Decoding + Tree Attention
Why Serving 1,000 Fine-Tunes Doesn't Need 1,000 GPUs

Why Serving 1,000 Fine-Tunes Doesn't Need 1,000 GPUs
LLM Distillation's Hidden Problem: When Teacher & Student Tokenizers Don't Match

LLM Distillation's Hidden Problem: When Teacher & Student Tokenizers Don't Match
Why Modern

Why Modern "Path Traced" Games Are 99% Hallucinated Rays
How Context Parallelism Splits Long Sequences Across GPUs

How Context Parallelism Splits Long Sequences Across GPUs