Run Ops Mlir Snippets

Run ops.mlir snippets (compile + execute)

Given an ops.mlir-style file (a module containing one func.func per unique TTIR op configuration), compile each function to TTMetal (or TTNN) and optionally execute on device.

The input can be:

• A single .mlir file (e.g. ops.mlir)
• A directory of .mlir files -- processes every *.mlir in it. Each file gets its own report. Caution: only point a directory at folders that contain ops-style snippet files, not raw/preprocessed model IR.

The driver script is tools/scripts/model_breakdown/run_ops_mlir_snippets.py.

Prerequisites

source env/activate
ttrt query --save-artifacts                    # creates system descriptor
export SYSTEM_DESC_PATH="$(pwd)/ttrt-artifacts/system_desc.ttsys"

Basic usage

Single file:

python tools/scripts/model_breakdown/run_ops_mlir_snippets.py path/to/ops.mlir

Multiple files (per-file reports + combined report at common parent):

python tools/scripts/model_breakdown/run_ops_mlir_snippets.py path/to/*/ops.mlir

Directory (processes every *.mlir in the dir):

python tools/scripts/model_breakdown/run_ops_mlir_snippets.py path/to/dir/

This compiles and executes every snippet. Each function is wrapped in its own module, compiled via compile_ttir_module_to_flatbuffer, and run with execute_fb. In directory mode, the device is opened once and shared across all files.

Flags

Common workflows

Compile-only triage (no device needed)

Use --skip-exec to find which ops fail at compile time without requiring hardware:

python tools/scripts/model_breakdown/run_ops_mlir_snippets.py path/to/ops.mlir --skip-exec

List or run one snippet

Use --list to see the functions in an ops.mlir, and combine --func with --skip-exec to compile one matching snippet without opening a device:

python tools/scripts/model_breakdown/run_ops_mlir_snippets.py path/to/ops.mlir --list
python tools/scripts/model_breakdown/run_ops_mlir_snippets.py path/to/ops.mlir --func add_0 --skip-exec

Multi-graph model directory

After running the ttir-model-op-analysis skill on a multi-graph directory like vllm_opt/, each graph gets its own subdirectory with an ops.mlir. Pass all of them in one command:

python tools/scripts/model_breakdown/run_ops_mlir_snippets.py vllm_opt/*/ops.mlir --skip-exec

This writes ops-run-report.txt next to each ops.mlir, plus a combined ops-run-report.txt at the common parent with per-file summaries and all failures in one place:

vllm_opt/
  ops-run-report.txt            # combined report across all graphs
  graph1/
    ops.mlir
    ops-run-report.txt          # compile results for graph1
  graph2/
    ops.mlir
    ops-run-report.txt          # compile results for graph2

Important: pass the specific ops.mlir files, not the subdirectories. The subdirectories also contain preprocessed.mlir (the full model graph), which is not a snippet file and will produce a useless failure report if the runner tries to process it.

Fail-fast to find the first broken op

python tools/scripts/model_breakdown/run_ops_mlir_snippets.py path/to/ops.mlir --fail-fast

Save artifacts for debugging

python tools/scripts/model_breakdown/run_ops_mlir_snippets.py path/to/ops.mlir \
    --save-artifacts --output-root /tmp/snippets --print-ir

Artifacts land in <output-root>/ops_mlir_snippets/<filename>/<func_name>/<target>/.

Report

The script writes a <stem>-run-report.txt in the same directory as each input .mlir file (e.g. ops-run-report.txt for ops.mlir). In directory mode, each file gets its own report. The report has three sections:

1. Summary at top -- target, mode, pass/fail counts at a glance.
2. Per-op table -- one row per function showing compile (and execute) status.
3. Failure details -- numbered list with the Python exception and the captured MLIR diagnostics (L1 memory exceeded, missing parser, etc.).

Example (compile-only):

target:  ttmetal
input:   /path/to/ops.mlir
mode:    compile-only
total:   50 ops

  compile: 47/50 passed, 3 failed

────────────────────────────────────────────────────────────────────────

  func_name   compile
  ──────────  ───────
  softmax_0   ok
  matmul_0    FAILED
  reshape_0   FAILED
  ...

────────────────────────────────────────────────────────────────────────

  Failure details (3)

  [1] matmul_0 — compile FAILED
      exception: Failed to run pass manager
      diagnostics:
        can't find feasible allocation because all 8 var(s) are bound
        error: 'func.func' op required L1 memory usage 3309568 exceeds
               memory capacity 1395424 (usable space is [103712, 1499136))

  [2] reshape_0 — compile FAILED
      exception: No parser found for opview <class '...ReshapeOp'>

With --skip-exec, the execute column is omitted. Diagnostics are captured from C-level stderr so MLIR allocator errors, verification failures, etc. appear in the report even though the Python exception only says "Failed to run pass manager".

Interpreting stdout

The script also prints a banner per snippet to stdout:

============================================================
Snippet: ops.mlir/softmax_0
============================================================
  compile: ok
  execute: ok

On failure you'll see compile: FAILED: <error> or execute: FAILED: <error>. At the end: either all N snippet(s) succeeded across M file(s) or N snippet(s) failed across M file(s).

Error handling

• Compile failures skip to the next snippet (unless --fail-fast).
• Execution failures close and re-open the device before continuing, so one hang doesn't block the rest of the run.
• If a snippet causes a device hang that persists across re-open, use --skip-exec to isolate compile issues, then test individual snippets by extracting the function into its own file.

Generating ops.mlir

If you don't already have an ops.mlir, see the ttir-model-op-analysis skill which produces one from a model's TTIR dump via tools/scripts/model_breakdown/ttir_model_op_inventory.py.