[UX][MetaSchedule] Tuning Interface

[junrushao]

Relax-related

• S2
• S4: nop

TIR-related

• S5
• S6

MetaSchedule-related

• S1
• S3
• S7
• S8
• S9

[Fixed] S1. Tune APIs return Database

Fixed: apache/tvm#12895

The tuning interface should return a ms.Database consistently across TIR and Relax, which could be further applied to Relax compilation pipeline.

[Fixed] S2. ApplyHistoryBest as a context class, but with a potentially better name

Fixed: apache/tvm#12520

with ApplyHistoryBest(database):
  artifact = relax.build(...)

Update: ApplyHistoryBest is removed with respect to making Database with-able:

with database:
  artifact = relax.build(...)

[Fixed] S3. Database.query returns tir.Schedule

Fixed: apache/tvm#12520

Currently it returns IRModule, which is less convenient to deal with when we want to inspect the trace - changing to returning tir.Schedule so that the trace is available.

[NOP] S4. Build a Relax pass to apply database to IRModule

I believe it's already there. CC: @tqchen @sunggg @YuchenJin.

[Fixed] S5. Trace printing should be wrapped with a function.

Fixed: apache/tvm#12507

Current:

>>> print(sch.trace)
b1 = sch.get_block(name="xxx")

Updated:

>>> print(sch.trace)
def schedule(sch):
  b1 = sch.get_block(name="xxx")

[Fixed] S6. Rendering trace with pigment

Fixed: apache/tvm#12507

Adopting the contribution from @ganler.

[Fixed] S7. Create database via work_dir with some auto-detect functionality

Fixed: apache/tvm#12643

Adding API: ms.Database.from_work_dir("/path/to/work_dir/").

Update: it's possible now with:

ms.database("json", work_dir="/path/to/work_dir")

[sunggg]

Thank you for the great suggestion!
Regarding S4, yes. We introduced ApplyHistoryBest as a pass and currently we are reviewing a PR for MS Tuning pass.

[junrushao]

S8. Handle it when excessive errors pop out

Stop tuning or fall back to debug mode

[junrushao]

S9. Use better table format when tuning in Jupyter