Skip to content

Fix packed INT8 resident-byte validation #68

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Alex-Wengg merged 1 commit into from
Sep 15, 2026
Merged

Fix packed INT8 resident-byte validation #68

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 12 additions & 2 deletions packages/acestep/src/webgpu/dit-graph.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,6 +52,8 @@ import {
import {
ACE_OPT_0009_DIT_DENSE_KERNEL_SET_ID,
ACE_OPT_0009_DIT_DENSE_RUNTIME_PROFILE,
ACE_OPT_0091_DIT_INT8_LAYER_BYTES,
ACE_OPT_0091_DIT_INT8_RESIDENT_WEIGHT_BYTES,
ACE_OPT_0091_DIT_INT8_RUNTIME_PROFILE,
ACE_OPT_0037_DIT_K4_LAYER_BYTES,
ACE_OPT_0037_DIT_K4_KERNEL_SET_ID,
Expand Down Expand Up @@ -554,11 +556,19 @@ export class AceDitResidentModel implements AceDitGraphModel {
densePhase,
denseRuntimeProfile,
);
const int8 =
denseRuntimeProfile === ACE_OPT_0091_DIT_INT8_RUNTIME_PROFILE;
const expectedDenseBytes = int8
? ACE_OPT_0091_DIT_INT8_LAYER_BYTES
: ACE_OPT_0037_DIT_K4_LAYER_BYTES;
const expectedResidentBytes = int8
? ACE_OPT_0091_DIT_INT8_RESIDENT_WEIGHT_BYTES
: ACE_OPT_0037_DIT_K4_RESIDENT_WEIGHT_BYTES;
if (
densePhase !== undefined &&
(phase.residentBytes !== ACE_REFERENCE_DIT_SHARED_WEIGHT_BYTES ||
densePhase.residentBytes !== ACE_OPT_0037_DIT_K4_LAYER_BYTES ||
this.residentBytes !== ACE_OPT_0037_DIT_K4_RESIDENT_WEIGHT_BYTES)
densePhase.residentBytes !== expectedDenseBytes ||
this.residentBytes !== expectedResidentBytes)
) {
throw new Error(
"ACE mixed DiT physical resident-byte replacement contract changed",
Expand Down
56 changes: 50 additions & 6 deletions packages/acestep/test/dit-backend.test.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,6 +48,9 @@ import {
import {
ACE_OPT_0009_DIT_DENSE_RUNTIME_PROFILE,
ACE_OPT_0009_DIT_MIXED_LAYER_BYTES,
ACE_OPT_0091_DIT_INT8_LAYER_BYTES,
ACE_OPT_0091_DIT_INT8_RESIDENT_WEIGHT_BYTES,
ACE_OPT_0091_DIT_INT8_RUNTIME_PROFILE,
ACE_OPT_0037_DIT_K4_RUNTIME_PROFILE,
ACE_OPT_0056_DIT_SELECTIVE_K4_RUNTIME_PROFILE,
ACE_REFERENCE_DIT_SHARED_WEIGHT_BYTES,
Expand Down Expand Up @@ -77,6 +80,8 @@ import {
ACE_DIT_GEMM_TILE_LAYOUT,
ACE_DIT_DENSE_FP16_TILE_LAYOUT,
ACE_DIT_DENSE_FP16_TRANSFORMATION,
ACE_DIT_DENSE_INT8_TILE_LAYOUT,
ACE_DIT_DENSE_INT8_TRANSFORMATION,
ACE_DIT_DENSE_K4_FP16_LAYOUT,
ACE_DIT_DENSE_K4_FP16_TRANSFORMATION,
} from "../src/model/manifest.js";
Expand Down Expand Up @@ -521,6 +526,23 @@ describe("ACE DiT resident phase ownership", () => {
);
expect(destroy).toHaveBeenCalledTimes(1);
});

it("accepts the packed INT8 resident-byte contract", () => {
const phases = fakeMixedDitPhases("opt-0091-int8");
const model = AceDitResidentModel.takeMixed(
phases.reference.value,
phases.mixed.value,
"reference-bf16",
ACE_OPT_0091_DIT_INT8_RUNTIME_PROFILE,
);

expect(model.residentBytes).toBe(
ACE_OPT_0091_DIT_INT8_RESIDENT_WEIGHT_BYTES,
);
model.destroy();
expect(phases.reference.destroy).toHaveBeenCalledOnce();
expect(phases.mixed.destroy).toHaveBeenCalledOnce();
});
});

describe("ACE DiT concrete factory", () => {
Expand Down Expand Up @@ -1533,7 +1555,8 @@ function overlap(left: readonly Lifetime[], right: readonly Lifetime[]): boolean
}

function fakeMixedDitPhases(
denseProfile: "opt-0009" | "opt-0037-k4" = "opt-0009",
denseProfile: "opt-0009" | "opt-0037-k4" | "opt-0091-int8" =
"opt-0009",
): Readonly<{
reference: Readonly<{
value: AceGpuTensorPhase;
Expand All @@ -1551,7 +1574,8 @@ function fakeMixedDitPhases(

function fakeDitPhase(
profile: "reference-shared" | "mixed-layers",
denseProfile: "opt-0009" | "opt-0037-k4" = "opt-0009",
denseProfile: "opt-0009" | "opt-0037-k4" | "opt-0091-int8" =
"opt-0009",
): Readonly<{
value: AceGpuTensorPhase;
destroy: ReturnType<typeof vi.fn>;
Expand All @@ -1567,11 +1591,15 @@ function fakeDitPhase(
const storageShape = dense
? denseProfile === "opt-0037-k4"
? [logicalShape[0]! / 128, logicalShape[1]! / 4, 4, 32, 4]
: denseProfile === "opt-0091-int8"
? [logicalShape[0]! / 256, logicalShape[1]! / 32, 2_176]
: [...logicalShape]
: [Math.ceil(elements / 2)];
const transformation = dense
? denseProfile === "opt-0037-k4"
? ACE_DIT_DENSE_K4_FP16_TRANSFORMATION
: denseProfile === "opt-0091-int8"
? ACE_DIT_DENSE_INT8_TRANSFORMATION
: ACE_DIT_DENSE_FP16_TRANSFORMATION
: tiled
? ACE_DIT_GEMM_PACKED_BF16_TRANSFORMATION
Expand All @@ -1585,14 +1613,26 @@ function fakeDitPhase(
logicalTensor: name,
logicalShape,
storageShape,
byteLength: elements * 2,
byteLength:
dense && denseProfile === "opt-0091-int8"
? (logicalShape[0]! / 256) *
(logicalShape[1]! / 32) *
8_704
: elements * 2,
phase: "dit",
lifetime: "dit",
source: `ace-turbo-weights:${name.slice("ace.".length)}`,
dtype: dense ? "float16" : "uint32-bf16-pairs",
dtype:
dense && denseProfile === "opt-0091-int8"
? "uint32-int8-fp16-blocks"
: dense
? "float16"
: "uint32-bf16-pairs",
layout: dense
? denseProfile === "opt-0037-k4"
? ACE_DIT_DENSE_K4_FP16_LAYOUT
: denseProfile === "opt-0091-int8"
? ACE_DIT_DENSE_INT8_TILE_LAYOUT
: ACE_DIT_DENSE_FP16_TILE_LAYOUT
: tiled
? ACE_DIT_GEMM_TILE_LAYOUT
Expand All @@ -1611,11 +1651,15 @@ function fakeDitPhase(
phases: Object.freeze(["dit"]),
residentBytes: profile === "reference-shared"
? ACE_REFERENCE_DIT_SHARED_WEIGHT_BYTES
: ACE_OPT_0009_DIT_MIXED_LAYER_BYTES,
: denseProfile === "opt-0091-int8"
? ACE_OPT_0091_DIT_INT8_LAYER_BYTES
: ACE_OPT_0009_DIT_MIXED_LAYER_BYTES,
packageManifest: {
profile: profile === "reference-shared"
? "reference"
: "fp16-dit-dense-experimental",
: denseProfile === "opt-0091-int8"
? "int8-dit-dense-experimental"
: "fp16-dit-dense-experimental",
},
logicalTensor,
binding: (name: string) => logicalTensor(name).parts[0]!.binding,
Expand Down
Loading