[XPU] Enable sequence parallel support for XPU (#38608)

Signed-off-by: chaojun-zhang <chaojun.zhang@intel.com>
Signed-off-by: Chaojun Zhang <chaojun.zhang@intel.com>
Signed-off-by: Chaojun,Zhang <chaojun.zhang@intel.com>
This commit is contained in:
Chaojun Zhang
2026-06-14 19:26:46 -07:00
committed by GitHub
parent 1801fad0ba
commit 2725c84aae
5 changed files with 126 additions and 17 deletions
@@ -72,24 +72,27 @@ def get_sequence_parallelism_threshold(
"""
from vllm.platforms import current_platform
if not current_platform.is_cuda():
return None
if current_platform.is_xpu():
min_hidden_size = 4096
min_per_gpu_size_mb = 8.0
elif current_platform.is_cuda():
capability = current_platform.get_device_capability()
if capability is None:
return None
capability = current_platform.get_device_capability()
if capability is None:
return None
# Collapse Blackwell variants (sm100/sm103/...) into one policy bucket.
if current_platform.is_device_capability_family(100):
device_capability = 100
else:
device_capability = capability.to_int()
# Collapse Blackwell variants (sm100/sm103/...) into one policy bucket.
if current_platform.is_device_capability_family(100):
device_capability = 100
# Check if device has configured thresholds
_hidden = SP_MIN_HIDDEN_SIZE.get(device_capability)
_gpu_mb = SP_MIN_PER_GPU_SIZE_MB.get(device_capability)
if _hidden is None or _gpu_mb is None:
return None
min_hidden_size, min_per_gpu_size_mb = _hidden, _gpu_mb
else:
device_capability = capability.to_int()
# Check if device has configured thresholds
min_hidden_size = SP_MIN_HIDDEN_SIZE.get(device_capability)
min_per_gpu_size_mb = SP_MIN_PER_GPU_SIZE_MB.get(device_capability)
if min_hidden_size is None or min_per_gpu_size_mb is None:
return None
# Only apply sequence parallelism for models meeting the size threshold
+3 -1
View File
@@ -29,6 +29,9 @@ if rocm_aiter_ops.is_enabled():
RocmAiterTritonAddRMSNormPadFusionPass,
)
if current_platform.is_cuda_alike() or current_platform.is_xpu():
from .fusion.sequence_parallelism import SequenceParallelismPass
if current_platform.is_cuda_alike():
from .fusion.act_quant_fusion import ActivationQuantFusionPass
from .fusion.attn_quant_fusion import AttnQuantFusionPass
@@ -37,7 +40,6 @@ if current_platform.is_cuda_alike():
from .fusion.qk_norm_rope_fusion import QKNormRoPEFusionPass
from .fusion.rms_quant_fusion import RMSNormQuantFusionPass
from .fusion.rope_kvcache_fusion import RopeKVCacheFusionPass
from .fusion.sequence_parallelism import SequenceParallelismPass
from .utility.scatter_split_replace import ScatterSplitReplacementPass
from .utility.split_coalescing import SplitCoalescingPass