[Bugfix][CI] Update CuTeDSL indexer Q sentinel test for migrated wrappers (#56594)
Signed-off-by: LopezCastroRoberto <rocastro@redhat.com>
Roberto L. Castro committed
Sep 12, 2026 at 18:12 UTC
c377114636db645e13594e66a6c4f4f0084adac2
2 files changed
+20
-20
tests/kernels/test_fused_indexer_q_rope_quant.py
+19
-19
@@ -289,16 +289,16 @@ def test_cutedsl_indexer_q_writes_stay_within_num_tokens(use_fp4, n_head):
289
dtype=torch.uint8,
290
device=device,
291
)
292
- mod.fused_indexer_q_rope_quant_mxfp4_cutedsl(
293
- positions[:num_tokens],
294
- q[:num_tokens],
295
- cos_sin_cache,
296
- weights[:num_tokens],
297
- 1.0,
298
- 1.0,
299
- outputs["q_packed"][:num_tokens],
300
- outputs["q_scale"][:num_tokens],
301
- weights_out[:num_tokens],
292
+ mod._INDEXER_Q_MXFP4_KERNEL(
293
+ positions=positions[:num_tokens],
294
+ q=q[:num_tokens],
295
+ cos_sin_cache=cos_sin_cache,
296
+ weights=weights[:num_tokens],
297
+ weights_softmax_scale=1.0,
298
+ weights_head_scale=1.0,
299
+ q_packed=outputs["q_packed"][:num_tokens],
300
+ q_scale=outputs["q_scale"][:num_tokens],
301
+ weights_out=weights_out[:num_tokens],
302
)
303
else:
304
outputs["q_fp8"] = torch.full(
@@ -307,15 +307,15 @@ def test_cutedsl_indexer_q_writes_stay_within_num_tokens(use_fp4, n_head):
307
dtype=torch.uint8,
308
device=device,
309
)
310
- mod.fused_indexer_q_rope_quant_fp8_cutedsl(
311
- positions[:num_tokens],
312
- q[:num_tokens],
313
- cos_sin_cache,
314
- weights[:num_tokens],
315
- 1.0,
316
- 1.0,
317
- outputs["q_fp8"][:num_tokens].view(torch.float8_e4m3fn),
318
- weights_out[:num_tokens],
310
+ mod._INDEXER_Q_FP8_KERNEL(
311
+ positions=positions[:num_tokens],
312
+ q=q[:num_tokens],
313
+ cos_sin_cache=cos_sin_cache,
314
+ weights=weights[:num_tokens],
315
+ weights_softmax_scale=1.0,
316
+ weights_head_scale=1.0,
317
+ q_fp8=outputs["q_fp8"][:num_tokens].view(torch.float8_e4m3fn),
318
+ weights_out=weights_out[:num_tokens],
319
)
320
torch.accelerator.synchronize()
321
vllm/models/deepseek_v4/nvidia/ops/fused_indexer_q_cutedsl.py
+1
-1
@@ -684,7 +684,7 @@ class IndexerQFp8Kernel(VllmCuTeDSLJitKernel["IndexerQFp8Kernel.CompileKey"]):
684
q,
685
cos_sin_cache,
686
weights,
687
- q_fp8,
687
+ q_fp8.view(torch.uint8),
688
weights_out,
689
float(weights_softmax_scale * weights_head_scale),
690
)