Jatinkumar Parmar·Sep 10Following the Bottleneck: Optimizing MiniMax M3 on AMD Instinct MI355XThe MiniMax M3 day-0 post described the first working vLLM implementation: MiniMax Sparse Attention (MSA), multimodal inputs, reasoning and…
InTraceOptbyAbhijeet Pendyala, Phd·Jul 22TraceML vs torch.profiler vs cProfile: what each one costs to find the same bottleneckThe same input-bound ResNet-18 run, found three ways. TraceML runs live, at 2% overhead, and named the bottleneck in one line…
AlterSquare·May 45 Things That Break First When a Startup Scales Past $2M ARRTL;DR: Systems that worked at $500K start cracking at $2M ARR — and the failures are predictable. Fragile architecture, unmanaged tech…
Sunil Rawat·Jan 13Re-examining Frontier-Scale Training EfficiencyLatency, utilization, and the architectural crossover
