AMD-AGI
Official@amd-agi
Offers specialized diagnostic capabilities for MaxText training runs, focusing on Prometheus TSDB metrics, log triage, and performance bottleneck identification.
Agent Skills by AMD-AGI
Showing 24 vetted skills indexed across 2 GitHub repositories.
AMD-AGIroofline
Estimates per-kernel roofline headroom from GPU profile data to prioritize optimization targets.
AMD-AGIadd-expert-skill-to-geak
Scaffold, validate, and submit expert GPU kernel optimization recipes to GEAK.
AMD-AGItuning-hipblaslt
Tune hipBLASLt GEMM kernels by racing solutions and replaying winners by solution index.
AMD-AGItuning-ck
Races Composable Kernel GEMM instances with ckProfiler and deploys winners through aiter per-op tuners.
AMD-AGItuning-core
Guides the six-step loop for tuning GEMM and GPU ops on AMD Instinct accelerators.
AMD-AGIenv-setup
Audit and install GPU tuning tools in AMD ROCm serving containers.
AMD-AGItuning-triton
Tune Triton and Gluon GPU kernels on AMD Instinct using AMD-specific autotune knobs and noise-aware measurement.
AMD-AGItuning-hip
Tune hand-written HIP kernels on AMD Instinct GPUs using launch geometry and rocprofv3 profiling.
AMD-AGItuning-in-vllm
Tune fused-MoE and GEMM kernels inside a vLLM serving deployment on AMD Instinct GPUs.
AMD-AGItuning-in-sglang
Tune fused-MoE and GEMM kernels inside SGLang serving deployments on AMD Instinct GPUs.
AMD-AGItuning-flydsl
Tune FlyDSL kernel configs on AMD Instinct GPUs and race them against other backends.
AMD-AGItuning-aiter
Tune GEMM and fused ops through aiter's gradlib tuner across hipblaslt, triton, flydsl, asm, and torch backends on AMD Instinct GPUs.
AMD-AGIflydsl_fp8_gemm_playbook
Implements a fused FlyDSL fp8 blockscale GEMM core for AMD MI300X down-projection layers.
AMD-AGIflydsl_rewrite_quantized_moe
Rewrites quantized Triton fused-MoE GEMM kernels into FlyDSL for AMD MI300X GPUs.
AMD-AGIflydsl_prefill_moe_stage2_fp8partial
Stores grouped-MoE stage-2 down-proj partials in fp8 to halve HBM traffic on gfx950.
AMD-AGIgluon_authoring
Ports tuned Triton kernels to Gluon on AMD CDNA GPUs with verified layout and pipeline recovery.
AMD-AGIflydsl_fp8_blockscale_gemm
Rewrites CK fp8 blockscale GEMM kernels to FlyDSL on gfx950 with software fp32 scaling.
AMD-AGIflydsl_fused_attention_backward
Authors a fused multi-GEMM FlyDSL attention backward kernel for CDNA3/CDNA4 AMD GPUs.
AMD-AGIensure_flydsl
Builds or reuses a version-gated FlyDSL installation for AMD GPU kernel optimization workflows.
AMD-AGIflydsl_decode_moe_stage1_blkmap
Fuses paired same-expert sort-blocks into double-height tiles to halve MoE stage-1 weight HBM traffic.
AMD-AGImla_tilelang_to_triton
Ports MLA decode attention kernels from TileLang to Triton on gfx942 GPUs.
AMD-AGItsdb-diagnosis
Diagnose training incidents by querying per-job Prometheus TSDB metrics.
AMD-AGIjob-log-triage
Triage MaxText training job logs to identify status and failure modes.
AMD-AGIperformance-analysis
Analyze MaxText training runs to identify performance bottlenecks using TSDB, TraceLens, and IRLens.
Frequently Asked Questions About AMD-AGI
FAQPage SchemaWhat specific tasks can be performed using these diagnostic capabilities?▼
These capabilities enable the systematic diagnosis of training incidents through Prometheus TSDB metric queries, the triage of MaxText training logs to determine job status, and the identification of performance bottlenecks using TraceLens and IRLens diagnostic outputs.
Which engineering personas benefit from these diagnostic skills?▼
These skills are designed for machine learning infrastructure engineers, performance engineers, and site reliability engineers tasked with maintaining large-scale MaxText training clusters and resolving complex job failures.
What are the primary dependencies for executing these performance analyses?▼
Execution requires an active MaxText training environment, access to Prometheus TSDB instances for metric retrieval, and availability of TraceLens and IRLens diagnostic data generated during the training run.