Add a plugin mechanism for operator overrides - #743

Open

andrej wants to merge 19 commits into

ROCm:mainfrom

andrej:andrej/replace-gemm

Commits on Sep 18, 2026

gemma4e: ship the FLMGEMM artifacts and the instruction-stream loader
andrej
committed
6bd454d
plugins: let a shared library override individual model operations

andrej
and
claude
committed
65ad1dc
AutoModel: load operator plugins for every model, not just gemma4e

andrej
and
claude
committed
bdd05f5
plugins: drop the lm_head role, which no engine declares

andrej
and
claude
committed
b6caa17
flm_gemm: dequantize on the device, and make the weight source a mode

andrej
and
claude
committed
bddb4d0
plugins: a README that leads with what the example plugin does

andrej
and
claude
committed
31ba1ac
plugins: draw the chart from bench3.sh output instead of literals

andrej
and
claude
committed
c75fdb8
flm_gemm: rebuild the dequant streams against the simplified sequence

andrej
and
claude
committed
f0d2df2
flm_gemm: pick up the current GEMM, and refuse an incomplete artifact set

andrej
and
claude
committed
246654e
plugins: one README, not two

andrej
and
claude
committed
746a8b8
plugins: one copy of each tool, and a self-contained reproduction

andrej
and
claude
committed
6a1b0a6
plugins: name the configurations rather than rank them

andrej
and
claude
committed
eff1f10
add accuracy remark
andrej
committed
30a9183
plugins: rename the example to iron_gemm

andrej
and
claude
committed
3a15582
plugins: drop the registry API nothing calls, and one duplicated index

andrej
and
claude
committed
0f95075
plugins: let each model name its own operations

andrej
and
claude
committed
b2c62ca
iron_gemm: serve Gemma4 E4B

andrej
and
claude
committed
2d2eb2a
gemma4e: rebuild the engine against the rebased operator declarations

andrej
and
claude
committed
6d6dfa3
plugins: document the mechanism, and a plugin that is only a plugin

andrej
and
claude
committed
f8c2e49

Navigation Menu

Add a plugin mechanism for operator overrides - #743

Add a plugin mechanism for operator overrides#743
andrej wants to merge 19 commits into
ROCm:mainfrom
andrej:andrej/replace-gemm

Commits on Sep 18, 2026

gemma4e: ship the FLMGEMM artifacts and the instruction-stream loader

plugins: let a shared library override individual model operations

AutoModel: load operator plugins for every model, not just gemma4e

plugins: drop the lm_head role, which no engine declares

flm_gemm: dequantize on the device, and make the weight source a mode

plugins: a README that leads with what the example plugin does

plugins: draw the chart from bench3.sh output instead of literals

flm_gemm: rebuild the dequant streams against the simplified sequence

flm_gemm: pick up the current GEMM, and refuse an incomplete artifact set

plugins: one README, not two

plugins: one copy of each tool, and a self-contained reproduction

plugins: name the configurations rather than rank them

add accuracy remark

plugins: rename the example to iron_gemm

plugins: drop the registry API nothing calls, and one duplicated index

plugins: let each model name its own operations

iron_gemm: serve Gemma4 E4B

gemma4e: rebuild the engine against the rebased operator declarations

plugins: document the mechanism, and a plugin that is only a plugin