Running LLMs on AMD GPUs With ROCm: What Actually Works
HuggingFace's Optimum-AMD recipe and ROCm 7.2.4's vLLM fixes make the MI300X a serviceable inference target, but newer attention kernels and training still trail CUDA.
The archive · Page 4 of 7
Infrastructure & Runtime
The serving stack, network fabric, and cloud-account substrate beneath production AI, where every throughput claim collides with rebuild windows, egress invoices, and control-plane risk.
The archive
73–96 of 145 articles · Newest first
Cloudflare Meerkat: What Globally Distributed Consensus Costs at the Edge
Cloudflare's Meerkat runs QuePaxa consensus at the edge, so every write waits on a cross-region quorum. The write-latency tax suits control-plane state, not transactions.
Vercel CDN Now Honors External Origin Cache-Control: Audit Your Headers
Since April 6, 2026, new Vercel projects honor Cache-Control from external origins by default, so operators must audit rewrite headers or risk stale, unintended responses.
AI Found Real Bugs in Cloudflare's Circl Crypto Library
zkSecurity's AI agent found seven real bugs in Cloudflare's CIRCL crypto library, all fixed upstream, showing AI-assisted review is now a baseline layer for crypto code.
Pruning RAG Context: What to Cut Before the LLM Sees It
Pruning RAG context is a ranking decision. Reranking and compression keep only answer-changing chunks, shifting cost from the prompt onto retrieval and shrinking the cite set.
Cloudflare's x402 Gateway: What Per-Request API Billing Actually Needs
Cloudflare's Monetization Gateway prices resources per request in stablecoins via HTTP 402, shifting fee friction to callers and splitting pricing into metered and flat plans.
Doubao 2.1 Pro: What 180 Trillion Daily Tokens Means for Inference Infrastructure
Doubao 2.1 Pro ships at ¥6/¥30 per million tokens. The family's 180 trillion daily tokens reset what Western inference stacks must assume about price and capacity.
Every CUDA Kernel Pays a Launch Tax: The Host-to-Device Walkthrough
Every CUDA kernel pays a fixed driver-queue tax before its first FLOP runs. The fusion, graphs, and batching sold as bandwidth wins mostly hide the launch overhead.
Vercel Montreal Region: Audit Residency Before You Migrate
A Vercel Montreal region only earns its cost when a legal rule forces data to stay in Canada. For every other workload, the real work is a residency audit, not a migration.
GLM-5.2 on vLLM and Ascend: Open Weights Beyond NVIDIA
GLM-5.2 ships MIT-licensed with same-week serving recipes for NVIDIA vLLM and Huawei Ascend NPUs, breaking the open-weights-but-NVIDIA-only trade-off for self-hosters.
How Vercel Runs Its Own CDN in Front of Discourse: A Self-Dogfooding Case Study
Vercel fronts its own Discourse forum with its CDN, but the edge cache only serves anonymous reads. Logged-in pages fall to the origin by design.
Vercel Runtime Logs Surface CDN Cache Hits, Not the Eviction Cause
Vercel's Runtime Logs now show CDN cache key, tags, and revalidation reason per request, but the eviction cause stays hidden, forcing manual header inspection.
Multimodal Knowledge Graph RAG vs Vector RAG: What MKG-RAG-Bench Shows
MKG-RAG-Bench isolates retrieval in multimodal knowledge graph RAG and finds it is the bottleneck. Adding images and graph edges costs more without guaranteed accuracy.
Vercel Observability Now Tracks Redirects and Rewrites Beside Function Errors
Vercel added redirect and rewrite telemetry to Observability, putting per-route errors beside function errors and shifting routing incidents toward on-call engineers.
Cloudflare Workflows Saga Rollbacks: Compensating Actions in Serverless Orchestration
Cloudflare Workflows saga rollbacks move compensation into declared per-step undo handlers, exposing the idempotency assumption every saga platform quietly offloads.
Static Corpus RAG: The Bible Case for Separating Churn from Algorithm Complexity
Fixed-corpus RAG collapses re-chunking, freshness sync, and citation grounding. The Bible's book.chapter.verse scheme isolates churn-driven complexity from the algorithmic.
Vercel's KIKO Milano Black Friday Case Study: What the Scaling Claims Skip
Vercel's KIKO Milano case study has no QPS, cost, or cold-start data. Collapsing skips checkout paths, Fluid savings assume I/O-bound work. Run the numbers before the peak.
Vercel Postgres vs Neon vs Supabase: When the Bundled DB Wins
Vercel Postgres is a Neon resell now owned by Databricks. The bundled database trims setup friction for Next.js teams but stacks three vendors where most assume one.
Fine-Tuning a 20B LLM With RLHF on a 24GB GPU: What Fits
The 20B RLHF on 24GB claim is real but the VRAM cost is the duplicated reward model, not the LoRA adapters. Batch size, sequence length, and throughput pay the price.
Vercel Flat Rate CDN Beta: Break-Even Math for Spiky Workloads, Tax for the Rest
Vercel's Flat Rate CDN beta replaces per-GB egress with a fixed fee; spiky workloads win, steady low-traffic sites pay more, and the Cloudflare-fronting cost case weakens.
Where DeepSeek Weights Actually Run on Vercel's AI Gateway
DeepSeek hit 17% of tokens on Vercel's AI Gateway in May 2026, but no source confirms where the weights run. Jurisdiction, not model origin, is the new gate.
Vercel's Anti-Lock-In Pitch: What the Open-Source Bet Still Locks In
Vercel markets Next.js and the AI SDK as open source and portable, but the paid platform, from deploy previews to Fluid Compute, does not travel with the code.
Vercel Adds Tag-Based CDN Cache Invalidation: Surrogate Keys at the Edge
Vercel's January 2026 Vercel-Cache-Tag ship brings surrogate-key cache invalidation to every plan, moving the cache contract into application-owned tag strings.
GLM 5.2 Fast on Vercel AI Gateway: What Routing Through Wafer Actually Buys
Vercel AI Gateway reportedly routes GLM 5.2 Fast through Wafer, putting Zhipu's 754B-parameter MoE coding model one config entry from any Vercel app. Routing is unverified.
