[WIP] Add HPU support to vLLM v1 - cont.

vllm/attention/selector.py

@@ -159,6 +159,10 @@ def _cached_get_attn_backend(
         logger.info("Using HPUAttention backend.")
         from vllm.attention.backends.hpu_attn import HPUAttentionBackend
         return HPUAttentionBackend
+    elif backend == _Backend.HPU_ATTN_V1:
+        logger.info("Using HPUAttentionV1 backend.")
+        from vllm.v1.attention.backends.hpu_attn import HPUAttentionBackendV1
+        return HPUAttentionBackendV1
     elif backend == _Backend.PALLAS:
         logger.info("Using Pallas backend.")
         from vllm.attention.backends.pallas import PallasAttentionBackend

vllm/compilation/backends.py

@@ -11,6 +11,7 @@
 import vllm.envs as envs
 from vllm.config import CompilationConfig
 from vllm.logger import init_logger
+from vllm.platforms import current_platform
 from vllm.utils import weak_ref_tensors
 
 from .counter import compilation_counter
@@ -230,7 +231,7 @@ def __init__(
         compilation_configs: CompilationConfig,
     ):
         global global_graph_pool
-        if global_graph_pool is None:
+        if global_graph_pool is None and current_platform.is_cuda_alike():
             global_graph_pool = torch.cuda.graph_pool_handle()
 
         # TODO: in the future, if we want to use multiple

vllm/distributed/parallel_state.py

@@ -959,6 +959,7 @@ def init_distributed_environment(
         "world_size=%d rank=%d local_rank=%d "
         "distributed_init_method=%s backend=%s", world_size, rank, local_rank,
         distributed_init_method, backend)
+    print(distributed_init_method)
     if not torch.distributed.is_initialized():
         assert distributed_init_method is not None, (
             "distributed_init_method must be provided when initializing "

vllm/platforms/hpu.py

@@ -2,6 +2,8 @@
 
 import torch
 
+from vllm import envs
+
 from .interface import Platform, PlatformEnum, _Backend
 
 if TYPE_CHECKING:
@@ -36,12 +38,21 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
         parallel_config = vllm_config.parallel_config
         if parallel_config.worker_cls == "auto":
             if scheduler_config.is_multi_step:
+                if envs.VLLM_USE_V1:
+                    raise NotImplementedError
                 parallel_config.worker_cls = \
                     "vllm.worker.multi_step_hpu_worker.MultiStepHPUWorker"
             elif vllm_config.speculative_config:
+                if envs.VLLM_USE_V1:
+                    raise NotImplementedError
                 parallel_config.worker_cls = \
                     "vllm.spec_decode.spec_decode_worker.create_spec_worker"
                 parallel_config.sd_worker_cls = \
                     "vllm.worker.hpu_worker.HPUWorker"
             else:
-                parallel_config.worker_cls = "vllm.worker.hpu_worker.HPUWorker"
+                if envs.VLLM_USE_V1:
+                    parallel_config.worker_cls = \
+                        "vllm.v1.worker.hpu_worker.HPUWorker"
+                else:
+                    parallel_config.worker_cls = \
+                        "vllm.worker.hpu_worker.HPUWorker"