Skip to content

Commit cb1e46a

Browse files
fix(inference): swap known-issue precisions back (GB300 FP8, MI355X FP4) (#442)
#441 applied the precision swap to the wrong entries. The GB300 Dynamo TRT MTP accuracy issue (NVIDIA/srt-slurm#51) affects FP8, and the MI355X MoRI SGLang MTP issue (sgl-project/sglang#27194) affects FP4. Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
1 parent 85e1e41 commit cb1e46a

2 files changed

Lines changed: 13 additions & 13 deletions

File tree

packages/app/src/lib/known-issues.test.ts

Lines changed: 11 additions & 11 deletions
Original file line numberDiff line numberDiff line change
@@ -5,32 +5,32 @@ import { KNOWN_CONFIG_ISSUES, knownIssueCsvNote, matchKnownConfigIssues } from '
55
const DSR1 = 'DeepSeek-R1-0528';
66

77
describe('matchKnownConfigIssues', () => {
8-
it('matches the GB300 Dynamo TRT MTP entry for DeepSeek R1 FP4', () => {
8+
it('matches the GB300 Dynamo TRT MTP entry for DeepSeek R1 FP8', () => {
99
const issues = matchKnownConfigIssues(DSR1, [
10-
{ hwKey: 'gb300_dynamo-trt_mtp', precision: 'fp4' },
10+
{ hwKey: 'gb300_dynamo-trt_mtp', precision: 'fp8' },
1111
]);
1212
expect(issues).toHaveLength(1);
1313
expect(issues[0].url).toBe('https://github.com/NVIDIA/srt-slurm/issues/51');
1414
});
1515

16-
it('does not match GB300 Dynamo TRT MTP for non-FP4 precisions', () => {
16+
it('does not match GB300 Dynamo TRT MTP for non-FP8 precisions', () => {
1717
const issues = matchKnownConfigIssues(DSR1, [
18-
{ hwKey: 'gb300_dynamo-trt_mtp', precision: 'fp8' },
18+
{ hwKey: 'gb300_dynamo-trt_mtp', precision: 'fp4' },
1919
]);
2020
expect(issues).toHaveLength(0);
2121
});
2222

23-
it('matches the MI355X MoRI SGLang MTP entry for DeepSeek R1 FP8', () => {
23+
it('matches the MI355X MoRI SGLang MTP entry for DeepSeek R1 FP4', () => {
2424
const issues = matchKnownConfigIssues(DSR1, [
25-
{ hwKey: 'mi355x_mori-sglang_mtp', precision: 'fp8' },
25+
{ hwKey: 'mi355x_mori-sglang_mtp', precision: 'fp4' },
2626
]);
2727
expect(issues).toHaveLength(1);
2828
expect(issues[0].url).toBe('https://github.com/sgl-project/sglang/issues/27194');
2929
});
3030

31-
it('does not match MI355X MoRI SGLang MTP for non-FP8 precisions', () => {
31+
it('does not match MI355X MoRI SGLang MTP for non-FP4 precisions', () => {
3232
const issues = matchKnownConfigIssues(DSR1, [
33-
{ hwKey: 'mi355x_mori-sglang_mtp', precision: 'fp4' },
33+
{ hwKey: 'mi355x_mori-sglang_mtp', precision: 'fp8' },
3434
]);
3535
expect(issues).toHaveLength(0);
3636
});
@@ -54,9 +54,9 @@ describe('matchKnownConfigIssues', () => {
5454

5555
it('returns each issue at most once even with many matching points', () => {
5656
const issues = matchKnownConfigIssues(DSR1, [
57-
{ hwKey: 'gb300_dynamo-trt_mtp', precision: 'fp4' },
58-
{ hwKey: 'gb300_dynamo-trt_mtp', precision: 'fp4' },
59-
{ hwKey: 'mi355x_mori-sglang_mtp', precision: 'fp8' },
57+
{ hwKey: 'gb300_dynamo-trt_mtp', precision: 'fp8' },
58+
{ hwKey: 'gb300_dynamo-trt_mtp', precision: 'fp8' },
59+
{ hwKey: 'mi355x_mori-sglang_mtp', precision: 'fp4' },
6060
]);
6161
expect(issues).toHaveLength(2);
6262
});

packages/app/src/lib/known-issues.ts

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -29,7 +29,7 @@ export const KNOWN_CONFIG_ISSUES: KnownConfigIssue[] = [
2929
{
3030
hwKey: 'gb300_dynamo-trt_mtp',
3131
model: Model.DeepSeek_R1,
32-
precisions: ['fp4'],
32+
precisions: ['fp8'],
3333
summary: 'Accuracy issues',
3434
filed: 'Apr 21, 2026',
3535
url: 'https://github.com/NVIDIA/srt-slurm/issues/51',
@@ -38,7 +38,7 @@ export const KNOWN_CONFIG_ISSUES: KnownConfigIssue[] = [
3838
{
3939
hwKey: 'mi355x_mori-sglang_mtp',
4040
model: Model.DeepSeek_R1,
41-
precisions: ['fp8'],
41+
precisions: ['fp4'],
4242
summary: 'Accuracy issues',
4343
filed: 'Jun 4, 2026',
4444
url: 'https://github.com/sgl-project/sglang/issues/27194',

0 commit comments

Comments
 (0)