Pale Meridian · Nebula

One e-graph. Programs by context.

Qwen3-30B-A3B · BF16 · H100

One global optimum is the wrong object. Hardware × dtype × mode × users remain part of extraction.

TENSAT · 2021 ↗one cost contextgraph → EqSat → one extraction

Pale Meridianmany contextsone saturation → compatible local sections

Contextual extraction

one saturation · many restrictions

u1 / u2 / u4 selector · accepted baseline rates · matched u1 causal result below

sm90 · BF16 · users={0}

Saturateonce · shared e-graph

Restrict

Checklocal compatibility

GlueProgram + LLIR · v5

ExecuteAurora · H100

MEASURED · historical baseline239.40 tok/s

4.177 ms stepu1 · 916 launches · accepted

MEASURED · 916 → 483 launches

output tok/s · BF16 · c1

TensorRT-LLMvendor context162.60

vLLMvendor context214.89

PM authored · Amatched237.99

PM EqSat · Bmatched284.48

Vendor: workload shape + H100 class; different harness and card
A ↔ B: same model, input, oracle, runtime, clocks, and physical H100
Claim: EqSat + contextual extraction as one path; no saturation-only marginal

run order

A1238.04authored

B1284.42contextual EqSat

B2284.53contextual EqSat

A2237.93authored

Pooled +19.54% · arm drift <0.05%.

compiler-only · unflown

Greedy483

EqSat435

Declaration-order greedy consumes an overlap that EqSat preserves. The selected plan is −9.94% smaller; compiler-generated, not an H100 speedup.