[CP V1 Deprecation 4/5] Canonicalize prefill CP API names (#36229)

This commit is contained in:
Baizhou Zhang
2026-09-08 16:03:36 -07:00
committed by GitHub
parent 559c7fa75b
commit ed183d45ac
33 changed files with 143 additions and 173 deletions
@@ -28,7 +28,7 @@ from sglang.srt.layers.cp.utils import (
cp_gather_after_forward,
cp_shard_model_inputs,
get_cp_strategy,
is_cp_v2_active,
is_cp_active,
prepare_cp_forward,
)
from sglang.srt.layers.pooler import EmbeddingPoolerOutput
@@ -281,8 +281,8 @@ class EagerRunner(BaseRunner):
if not self.enable_pdmux:
forward_batch = self.load_batch(forward_batch, pp_proxy_tensors)
cp_v2_active = is_cp_v2_active(forward_batch)
if cp_v2_active:
cp_active = is_cp_active(forward_batch)
if cp_active:
prepare_cp_forward(forward_batch)
# Target verify can arrive with ``forward_metadata_ready`` set by an
@@ -293,7 +293,7 @@ class EagerRunner(BaseRunner):
# directly from the live ``spec_info`` tensors.
if (
forward_batch.needs_forward_metadata_init()
or cp_v2_active
or cp_active
or forward_batch.forward_mode.is_target_verify()
):
if model_runner.ps.attn_dcp_size > 1 and hasattr(
@@ -330,7 +330,7 @@ class EagerRunner(BaseRunner):
torch.get_device_module(model_runner.device),
)
if not cp_v2_active:
if not cp_active:
forward_batch.attn_cp_metadata = None
category = (
@@ -344,7 +344,7 @@ class EagerRunner(BaseRunner):
_is_hip
and pcg_runner is not None
and not isinstance(pcg_runner, EagerRunner)
and not cp_v2_active
and not cp_active
):
# HIP PCG eager fallback: enter the PCG context so Dynamo guards
# and PCG-specific MoE/attention paths stay consistent.
@@ -366,8 +366,8 @@ class EagerRunner(BaseRunner):
forward_batch,
**kwargs,
)
elif cp_v2_active:
ret = self._execute_extend_cp_v2(forward_batch, kwargs)
elif cp_active:
ret = self._execute_extend_cp(forward_batch, kwargs)
else:
ret = model_runner.model.forward(
forward_batch.input_ids,
@@ -377,7 +377,7 @@ class EagerRunner(BaseRunner):
)
return ret
def _execute_extend_cp_v2(
def _execute_extend_cp(
self, forward_batch: ForwardBatch, kwargs: dict
) -> Union[LogitsProcessorOutput, PPProxyTensors]:
"""CP extend: shard inputs at the model boundary, run the body on the