Shakeeb Alireza
Community@shakfu
Shakeeb Alireza's skill registry covers Mojo programming and Modular MAX model serving, importing, debugging, profiling, benchmarking, and evaluation on NVIDIA and AMD GPUs.
Agent Skills by Shakeeb Alireza
Showing 10 vetted skills indexed across 1 GitHub repositories.
shakfumojo-syntax
Generates Mojo code using current syntax, conventions, and removed-construct replacements.
shakfuimport-model
Ports Hugging Face model architectures into MAX with scaffold, implementation, and verification phases.
shakfudebug-model
Diagnose silent output corruption in MAX model ports using per-layer tensor-dump comparison against a reference.
shakfumojo-python-interop
Write Mojo code that calls Python libraries and exposes Mojo types as Python extension modules.
shakfuprofile-model
Profile MAX model inference to measure GPU utilization and identify dominant kernels.
shakfueval-model
Benchmarks text models served by MAX against standard datasets like GSM8K and MMLU.
shakfunew-modular-project
Initializes new Mojo or MAX projects with pixi, uv, pip, or conda environments.
shakfumojo-gpu-fundamentals
Guides writing Mojo GPU kernels for NVIDIA, AMD, and Apple accelerators.
shakfuserve-model
Launch and configure OpenAI-compatible model servers with MAX's max serve command.
shakfubenchmark-model
Benchmark MAX model servers to measure throughput, latency, and GPU utilization.
Frequently Asked Questions About Shakeeb Alireza
FAQPage SchemaWhat tasks can I accomplish with shakfu's Mojo and MAX skills?▼
You can write Mojo code with current syntax, program NVIDIA, AMD, and Apple GPUs, interoperate with Python, scaffold new Mojo/MAX projects, import Hugging Face models into MAX, serve them via max serve, debug output parity failures, profile GPU kernels, benchmark latency and throughput, and evaluate accuracy on GSM8K, MMLU, and similar benchmarks.
Who are these Mojo and MAX skills designed for?▼
They target ML engineers and systems developers working with the Modular stack: engineers porting Hugging Face architectures into MAX, debugging quantization or multi-GPU parity issues, writing Mojo GPU kernels, and performance engineers measuring inference throughput, TTFT, TPOT, and GPU saturation on served models.
What environment and dependencies do these skills require?▼
Most skills require a pixi or pip environment with MAX installed (often the nightly channel), network access to the Hugging Face Hub, and a GPU for serving, dumping, and verification. Kernel profiling additionally needs vendor profilers: nsys or ncu for NVIDIA, rocprofv3 for AMD.
How do I serve and benchmark a model using these skills?▼
Use serve-model to launch max serve against a Hugging Face repo or local checkpoint with correct flags like devices, quantization-encoding, and max-length. Then run benchmark-model against the endpoint to measure tokens-per-second, TTFT, TPOT, and GPU utilization, optionally collecting GPU stats on the same NVIDIA machine.
How do the skills handle a MAX model that serves but generates wrong output?▼
The debug-model skill addresses silent corruption: gibberish output, greedy token mismatches, or coherent-then-diverging generation. It mandates reference-vs-MAX tensor-dump comparators first, numeric verification of fixes before recompiling, and serve-vs-pipeline bisection when tensor dumps match but generated text still diverges.