github-actions[bot]
f9fcc7adab
date: 2026-06-28T09:48:35 master commit: da6313dbe95b3f24bb5d8018b0e5f950f5823ca7
62 lines
2.2 KiB
Plaintext
62 lines
2.2 KiB
Plaintext
VIZ is a tool for inspecting tinygrad's compilation process and performance profiling.
|
|
|
|
to use:
|
|
1. Run tinygrad with VIZ=1 (this saves the pkls and launches the server in interactive shells)
|
|
2. That's it!
|
|
|
|
This can:
|
|
1. See all schedules
|
|
2. See all graphs and how they were rewritten
|
|
3. See generated code
|
|
4. See profile
|
|
5. (AMD only) see instruction level SQTT profile
|
|
|
|
VIZ pkls can be viewed in two ways:
|
|
|
|
1. Web browser: python -m tinygrad.viz.serve
|
|
2. Command line: python -m tinygrad.viz.cli
|
|
|
|
By default, VIZ UIs automatically load the latest files.
|
|
|
|
user story: viewing profiling data
|
|
* tinygrad ran 32 LLM decode steps:
|
|
web: click "profiler", view the timeline of all python codegen and GPU kernels.
|
|
cli: Run `DEBUG=3 python -m tinygrad.viz.cli --json` to extract kernel timing info and ASTs in JSON format.
|
|
- note: Make sure to add NO_COLOR=1 to disable colored output.
|
|
|
|
user story: viewing code
|
|
* tinygrad ran 3 schedules: init the model + first train step, train step, test step
|
|
* schedule 1 (123) = main.py:97
|
|
* schedule 2 (97) = main.py:97
|
|
* schedule 3 (10) = main.py:145
|
|
* web: click "schedule 1", get list of kernels (like DEBUG=2)
|
|
* cli: `python -m tinygrad.viz.cli -s TINY "Schedule 3 Kernels n1"`
|
|
* kernel 1 "E_34_34" -- 'sin'
|
|
* kernel 2 "R_4545"
|
|
* web: click "E_34_34"
|
|
* cli: `python -m tinygrad.viz.cli -s TINY "do_to_program for E_34_34" "initial symbolic"`
|
|
* pre-rewritten UOp graph (step through rewrite here)
|
|
* post-rewritten UOp graph
|
|
* UOp list
|
|
* generated code
|
|
|
|
user story: debugging scheduler
|
|
* tinygrad ran 3 schedules: init the model + first train step, train step, test step
|
|
* ...
|
|
* click "schedule 1 graph", get a graph of the schedule in UOps
|
|
* step through rewrite rules
|
|
* see how things are broken into kernels
|
|
* see why two kernels didn't fuse
|
|
|
|
user story: SQTT / PMC profiling
|
|
|
|
note: SQTT has additional overhead, to enable it, set VIZ=2.
|
|
|
|
* tinygrad ran custom assembly GEMM kernel.
|
|
* web: click "SQTT gemm SE:1 PKTS", see wave instruction scheduling and CU execution unit occupancy at every clock cycle.
|
|
* cli: python -m tinygrad.viz.cli -s "kernel SQTT SE:0 PKTS"
|
|
|
|
* get bank conflicts:
|
|
* web: click "gemm PMC"
|
|
* cli: python -m tinygrad.viz.cli -s "gemm PMC" | rg -A 16 SQC_LDS_BANK_CONFLICT
|