A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[CI/Build] Fix and re-enable v1 PP test on CI (#25496)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Isotr0py committed
Sep 24, 2025 at 05:58 UTC
ae002924e96bd17cfc690c266623c340ff28a70f
4 files changed
-24
tests/distributed/test_pipeline_parallel.py
-3
@@ -382,7 +382,6 @@ def test_tp_language_generation(
382
test_options: PPTestOptions,
383
num_gpus_available,
384
):
385
- pytest.skip("Skipping the test until V1 passes it.")
385
_compare_tp(model_id,
386
parallel_setup,
387
distributed_backend,
@@ -410,7 +409,6 @@ def test_tp_language_embedding(
409
test_options: PPTestOptions,
410
num_gpus_available,
411
):
413
- pytest.skip("Skipping the test until V1 passes it.")
412
_compare_tp(model_id,
413
parallel_setup,
414
distributed_backend,
@@ -438,7 +436,6 @@ def test_tp_multimodal_generation(
436
test_options: PPTestOptions,
437
num_gpus_available,
438
):
441
- pytest.skip("Skipping the test until V1 passes it.")
439
_compare_tp(model_id,
440
parallel_setup,
441
distributed_backend,
vllm/model_executor/models/granite.py
-7
@@ -308,13 +308,11 @@ class GraniteModel(nn.Module):
308
hidden_states = inputs_embeds
309
else:
310
hidden_states = self.get_input_embeddings(input_ids)
311
- residual = None
311
312
hidden_states *= self.config.embedding_multiplier
313
else:
314
assert intermediate_tensors is not None
315
hidden_states = intermediate_tensors["hidden_states"]
317
- residual = intermediate_tensors["residual"]
316
317
for layer in islice(self.layers, self.start_layer, self.end_layer):
318
hidden_states = layer(positions, hidden_states)
@@ -322,7 +320,6 @@ class GraniteModel(nn.Module):
320
if not get_pp_group().is_last_rank:
321
return IntermediateTensors({
322
"hidden_states": hidden_states,
325
- "residual": residual
323
})
324
325
hidden_states = self.norm(hidden_states)
@@ -475,10 +472,6 @@ class GraniteForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
472
torch.zeros((batch_size, self.config.hidden_size),
473
dtype=dtype,
474
device=device),
478
- "residual":
479
- torch.zeros((batch_size, self.config.hidden_size),
480
- dtype=dtype,
481
- device=device),
475
})
476
477
def load_weights(self, weights: Iterable[tuple[str,
vllm/model_executor/models/granitemoe.py
-7
@@ -298,17 +298,14 @@ class GraniteMoeModel(nn.Module):
298
else:
299
hidden_states = self.get_input_embeddings(input_ids)
300
hidden_states *= self.embedding_multiplier
301
- residual = None
301
else:
302
assert intermediate_tensors is not None
303
hidden_states = intermediate_tensors["hidden_states"]
305
- residual = intermediate_tensors["residual"]
304
for layer in islice(self.layers, self.start_layer, self.end_layer):
305
hidden_states = layer(positions, hidden_states)
306
if not get_pp_group().is_last_rank:
307
return IntermediateTensors({
308
"hidden_states": hidden_states,
311
- "residual": residual
309
})
310
hidden_states = self.norm(hidden_states)
311
return hidden_states
@@ -523,10 +520,6 @@ class GraniteMoeForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
520
torch.zeros((batch_size, self.config.hidden_size),
521
dtype=dtype,
522
device=device),
526
- "residual":
527
- torch.zeros((batch_size, self.config.hidden_size),
528
- dtype=dtype,
529
- device=device),
523
})
524
525
def load_weights(self, weights: Iterable[tuple[str,
vllm/model_executor/models/granitemoeshared.py
-7
@@ -195,17 +195,14 @@ class GraniteMoeSharedModel(nn.Module):
195
else:
196
hidden_states = self.get_input_embeddings(input_ids)
197
hidden_states *= self.embedding_multiplier
198
- residual = None
198
else:
199
assert intermediate_tensors is not None
200
hidden_states = intermediate_tensors["hidden_states"]
202
- residual = intermediate_tensors["residual"]
201
for layer in islice(self.layers, self.start_layer, self.end_layer):
202
hidden_states = layer(positions, hidden_states)
203
if not get_pp_group().is_last_rank:
204
return IntermediateTensors({
205
"hidden_states": hidden_states,
208
- "residual": residual
206
})
207
hidden_states = self.norm(hidden_states)
208
return hidden_states
@@ -323,10 +320,6 @@ class GraniteMoeSharedForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
320
torch.zeros((batch_size, self.config.hidden_size),
321
dtype=dtype,
322
device=device),
326
- "residual":
327
- torch.zeros((batch_size, self.config.hidden_size),
328
- dtype=dtype,
329
- device=device),
323
})
324
325
def load_weights(self, weights: Iterable[tuple[str,