Refactor vllm_hpu -> vllm_gaudi

.gitignore

@@ -1,5 +1,5 @@
 # version file generated by setuptools-scm
-/vllm_hpu/_version.py
+/vllm_gaudi/_version.py
 
 # Byte-compiled / optimized / DLL files
 __pycache__/

docs/.nav.yml

@@ -33,5 +33,5 @@ nav:
   - API Reference:
     - Summary: api/README.md
     - Contents:
-      - glob: api/vllm_hpu/*
+      - glob: api/vllm_gaudi/*
         preserve_directory_names: true

docs/api/README.md

@@ -2,4 +2,4 @@
 
 [](){ #pkg_overview }
 ## Full package overview
-::: vllm_hpu
+::: vllm_gaudi

mkdocs.yaml

@@ -55,7 +55,7 @@ plugins:
   - awesome-nav
   # For API reference generation
   - api-autonav:
-      modules: ["vllm_hpu"]
+      modules: ["vllm_gaudi"]
       api_root_uri: "api"
   - mkdocstrings:
       handlers:

pyproject.toml

@@ -10,7 +10,7 @@ build-backend = "setuptools.build_meta"
 
 
 [project]
-name = "vllm_hpu"
+name = "vllm_gaudi"
 authors = [{name = "Intel"}]
 license = "Apache-2.0"
 readme = "README.md"
@@ -29,25 +29,25 @@ requires-python = ">=3.9,<3.13"
 dynamic = [ "version", "dependencies", "optional-dependencies"]
 
 [project.urls]
-Homepage = "https://github.com/HabanaAI/vllm-hpu-extension"
+Homepage = "https://github.com/vllm-project/vllm-gaudi"
 
 
 [tool.setuptools_scm]
 # no extra settings needed, presence enables setuptools-scm
 
 [tool.setuptools.packages.find]
 where = ["."]
-include = ["vllm_hpu"]
+include = ["vllm_gaudi"]
 
 [tool.yapfignore]
 ignore_patterns = [
     "build/**", 
-    "vllm_hpu/extension/**"  # NOTE(kzawora): re-enable this once extension refactor is ready
+    "vllm_gaudi/extension/**"  # NOTE(kzawora): re-enable this once extension refactor is ready
 ]
 
 [tool.ruff]
 # Allow lines to be as long as 80.
-extend-exclude = ["vllm_hpu/extension/**"]  # NOTE(kzawora): re-enable this once extension refactor is ready
+extend-exclude = ["vllm_gaudi/extension/**"]  # NOTE(kzawora): re-enable this once extension refactor is ready
 line-length = 80
 
 [tool.ruff.lint]
@@ -89,7 +89,7 @@ follow_imports = "silent"
 # After fixing type errors resulting from follow_imports: "skip" -> "silent",
 # move the directory here and remove it from tools/mypy.sh
 files = [
-    "vllm_hpu/*.py",
+    "vllm_gaudi/*.py",
 ]
 
 

setup.py

@@ -5,7 +5,7 @@
 from setuptools_scm import get_version
 
 try:
-    VERSION = get_version(write_to="vllm_hpu/_version.py")
+    VERSION = get_version(write_to="vllm_gaudi/_version.py")
 except LookupError:
     # The checkout action in github action CI does not checkout the tag. It
     # only checks out the commit. In this case, we set a dummy version.
@@ -39,12 +39,12 @@ def _read_requirements(filename: str) -> list[str]:
     try:
         requirements = _read_requirements("requirements.txt")
     except ValueError:
-        print("Failed to read requirements.txt in vllm_hpu.")
+        print("Failed to read requirements.txt in vllm_gaudi.")
     return requirements
 
 
 setup(
-    name="vllm_hpu",
+    name="vllm_gaudi",
     version=VERSION,
     author="Intel",
     long_description="Intel Gaudi plugin package for vLLM.",
@@ -63,7 +63,7 @@ def _read_requirements(filename: str) -> list[str]:
     ext_modules=ext_modules,
     extras_require={},
     entry_points={
-        "vllm.platform_plugins": ["hpu = vllm_hpu:register"],
-        "vllm.general_plugins": ["hpu_custom_ops = vllm_hpu:register_ops"],
+        "vllm.platform_plugins": ["hpu = vllm_gaudi:register"],
+        "vllm.general_plugins": ["hpu_custom_ops = vllm_gaudi:register_ops"],
     },
 )

tools/mypy.sh

@@ -22,8 +22,8 @@ run_mypy() {
 
 run_mypy # Note that this is less strict than CI
 run_mypy tests
-run_mypy vllm_hpu/attention
-run_mypy vllm_hpu/distributed
-#run_mypy vllm_hpu/extension # NOTE(kzawora): re-enable this once extension refactor is ready
-run_mypy vllm_hpu/ops
-run_mypy vllm_hpu/v1
+run_mypy vllm_gaudi/attention
+run_mypy vllm_gaudi/distributed
+#run_mypy vllm_gaudi/extension # NOTE(kzawora): re-enable this once extension refactor is ready
+run_mypy vllm_gaudi/ops
+run_mypy vllm_gaudi/v1

vllm_hpu/__init__.py → vllm_gaudi/__init__.py

@@ -1,4 +1,4 @@
-from vllm_hpu.platform import HpuPlatform
+from vllm_gaudi.platform import HpuPlatform
 import os
 
 
@@ -8,9 +8,9 @@ def register():
     if os.getenv("VLLM_WEIGHT_LOAD_FORCE_SYNC",
                  "false").lower() in ("true", "1"):
         HpuPlatform.set_synchronized_weight_loader()
-    return "vllm_hpu.platform.HpuPlatform"
+    return "vllm_gaudi.platform.HpuPlatform"
 
 
 def register_ops():
     """Register custom ops for the HPU platform."""
-    import vllm_hpu.ops  # noqa: F401
+    import vllm_gaudi.ops  # noqa: F401

vllm_hpu/attention/backends/hpu_attn.py → vllm_gaudi/attention/backends/hpu_attn.py

@@ -9,20 +9,20 @@
 from typing import Any, Optional
 
 import torch
-import vllm_hpu.extension.kernels as kernels
-import vllm_hpu.extension.ops as ops
-from vllm_hpu.extension.runtime import get_config
-from vllm_hpu.extension.utils import (FP8Matmul, Matmul, ModuleFusedSDPA,
-                                      Softmax, VLLMFP8KVCache, VLLMKVCache)
+import vllm_gaudi.extension.kernels as kernels
+import vllm_gaudi.extension.ops as ops
+from vllm_gaudi.extension.runtime import get_config
+from vllm_gaudi.extension.utils import (FP8Matmul, Matmul, ModuleFusedSDPA,
+                                        Softmax, VLLMFP8KVCache, VLLMKVCache)
 
 from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
                                               AttentionLayer,
                                               AttentionMetadata, AttentionType)
 
 from vllm.attention.backends.mla.common import MLACommonImpl
 from vllm.attention.backends.utils import CommonAttentionState
-from vllm_hpu.attention.ops.hpu_paged_attn import (HPUPagedAttention,
-                                                   HPUPagedAttentionMetadata)
+from vllm_gaudi.attention.ops.hpu_paged_attn import (HPUPagedAttention,
+                                                     HPUPagedAttentionMetadata)
 from vllm.logger import init_logger
 
 logger = init_logger(__name__)

vllm_hpu/attention/ops/hpu_paged_attn.py → vllm_gaudi/attention/ops/hpu_paged_attn.py

@@ -8,7 +8,7 @@
 from typing import Optional
 
 import torch
-from vllm_hpu.extension import cache_ops, ops
+from vllm_gaudi.extension import cache_ops, ops
 
 # Should be the same as PARTITION_SIZE in `paged_attention_v2_launcher`.
 _PARTITION_SIZE = 512

vllm_hpu/extension/bucketing/common.py → vllm_gaudi/extension/bucketing/common.py

@@ -42,9 +42,9 @@ def get_bucketing_context():
     use_exponential_bucketing = os.environ.get(
         'VLLM_EXPONENTIAL_BUCKETING', 'true').lower() == 'true'
     if use_exponential_bucketing:
-        from vllm_hpu.extension.bucketing.exponential import (
+        from vllm_gaudi.extension.bucketing.exponential import (
             HPUExponentialBucketingContext as HPUBucketingContext)
     else:
-        from vllm_hpu.extension.bucketing.linear import HPUBucketingContext
+        from vllm_gaudi.extension.bucketing.linear import HPUBucketingContext
 
     return HPUBucketingContext

vllm_hpu/extension/bucketing/exponential.py → vllm_gaudi/extension/bucketing/exponential.py

@@ -7,7 +7,7 @@
 from typing import Set, Tuple
 from .common import WeakSingleton
 
-from vllm_hpu.extension.runtime import get_config
+from vllm_gaudi.extension.runtime import get_config
 
 logger = logging.getLogger(__name__)
 

vllm_hpu/extension/bucketing/linear.py → vllm_gaudi/extension/bucketing/linear.py

@@ -6,7 +6,7 @@
 from typing import Tuple
 from .common import WeakSingleton
 
-from vllm_hpu.extension.runtime import get_config
+from vllm_gaudi.extension.runtime import get_config
 
 logger = logging.getLogger(__name__)
 

vllm_hpu/extension/environment.py → vllm_gaudi/extension/environment.py

@@ -20,7 +20,7 @@ def _get_hw(_):
             return "gaudi2"
         case htexp.synDeviceType.synDeviceGaudi3:
             return "gaudi3"
-    from vllm_hpu.extension.utils import is_fake_hpu
+    from vllm_gaudi.extension.utils import is_fake_hpu
     if is_fake_hpu():
         return "cpu"
     logger().warning(f'Unknown device type: {device_type}')
@@ -40,7 +40,7 @@ def _get_build(_):
     if output.returncode == 0 and match:
         return match.group('version')
     # In cpu-test environment we don't have access to habana-torch-plugin
-    from vllm_hpu.extension.utils import is_fake_hpu
+    from vllm_gaudi.extension.utils import is_fake_hpu
     result = '0.0.0.0' if is_fake_hpu() else None
     logger().warning(f"Unable to detect habana-torch-plugin version! Returning: {result}")
     return result
@@ -58,7 +58,7 @@ def set_vllm_config(cfg):
     # t.compile is very picky about what functions we can call inside modules
     # since this is the last step we can force recompilation of config to
     # ensure all values are computed before entering the model
-    from vllm_hpu.extension.runtime import get_config
+    from vllm_gaudi.extension.runtime import get_config
     get_config().finalize()
 
 

vllm_hpu/extension/features.py → vllm_gaudi/extension/features.py

@@ -5,11 +5,11 @@
 # LICENSE file in the root directory of this source tree.
 ###############################################################################
 
-from vllm_hpu.extension.config import (Not, Hardware, VersionRange, ModelType,
+from vllm_gaudi.extension.config import (Not, Hardware, VersionRange, ModelType,
                                        Kernel, FirstEnabled, All, Value, Env,
                                        Disabled, Engine, choice, boolean,
                                        to_dict, split_values_and_flags)
-from vllm_hpu.extension.kernels import fsdpa, block_softmax_adjustment
+from vllm_gaudi.extension.kernels import fsdpa, block_softmax_adjustment
 
 
 def get_user_flags():

vllm_hpu/extension/ops.py → vllm_gaudi/extension/ops.py

@@ -12,7 +12,7 @@
 import torch.nn.functional as F
 import math
 import habana_frameworks.torch.core as htcore
-from vllm_hpu.extension.runtime import get_config
+from vllm_gaudi.extension.runtime import get_config
 import habana_frameworks.torch.utils.experimental as htexp
 
 is_hpu_gaudi2 = htexp._get_device_type() == htexp.synDeviceType.synDeviceGaudi2

vllm_hpu/extension/profiler.py → vllm_gaudi/extension/profiler.py

@@ -15,7 +15,7 @@
 import uuid
 from habana_frameworks.torch import torch
 
-from vllm_hpu.extension.utils import is_fake_hpu
+from vllm_gaudi.extension.utils import is_fake_hpu
 from .logger import logger
 
 class FileWriter(threading.Thread):

vllm_hpu/extension/runtime.py → vllm_gaudi/extension/runtime.py

@@ -6,10 +6,10 @@
 ###############################################################################
 
 
-from vllm_hpu.extension.environment import get_environment
-from vllm_hpu.extension.features import get_features, get_user_flags, get_experimental_flags
-from vllm_hpu.extension.config import Config
-from vllm_hpu.extension.logger import logger
+from vllm_gaudi.extension.environment import get_environment
+from vllm_gaudi.extension.features import get_features, get_user_flags, get_experimental_flags
+from vllm_gaudi.extension.config import Config
+from vllm_gaudi.extension.logger import logger
 
 
 DETECTED = None

vllm_hpu/extension/test_bucketing.py → vllm_gaudi/extension/test_bucketing.py

@@ -8,8 +8,8 @@
 
 import pytest
 
-import vllm_hpu.extension.bucketing.linear as linear
-from vllm_hpu.extension.bucketing.exponential import HPUExponentialBucketingContext
+import vllm_gaudi.extension.bucketing.linear as linear
+from vllm_gaudi.extension.bucketing.exponential import HPUExponentialBucketingContext
 
 
 @pytest.fixture

vllm_hpu/extension/test_flags.py → vllm_gaudi/extension/test_flags.py

@@ -7,7 +7,7 @@
 
 import os
 import pytest
-from vllm_hpu.extension.config import VersionRange, Config, Kernel, Env, boolean, All, Not, Eq, Enabled, FirstEnabled, choice
+from vllm_gaudi.extension.config import VersionRange, Config, Kernel, Env, boolean, All, Not, Eq, Enabled, FirstEnabled, choice
 
 
 def with_cfg(fn):

vllm_hpu/extension/utils.py → vllm_gaudi/extension/utils.py

@@ -11,7 +11,7 @@
 import habana_frameworks.torch as htorch
 import torch
 
-from vllm_hpu.extension.runtime import get_config
+from vllm_gaudi.extension.runtime import get_config
 
 
 @lru_cache(maxsize=None)

vllm_gaudi/ops/__init__.py

@@ -0,0 +1,6 @@
+import vllm_gaudi.ops.hpu_fused_moe  # noqa
+import vllm_gaudi.ops.hpu_layernorm  # noqa
+import vllm_gaudi.ops.hpu_lora  # noqa
+import vllm_gaudi.ops.hpu_rotary_embedding  # noqa
+import vllm_gaudi.ops.hpu_compressed_tensors  # noqa
+import vllm_gaudi.ops.hpu_fp8  # noqa

vllm_hpu/ops/hpu_compressed_tensors.py → vllm_gaudi/ops/hpu_compressed_tensors.py

@@ -19,8 +19,8 @@
     compressed_tensors_moe)
 from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors_moe import (  # noqa: E501
     CompressedTensorsW8A8Fp8MoEMethod)
-import vllm_hpu.extension.ops as hpu_ops
-from vllm_hpu.extension.ops import VllmMixtureOfExpertsOpFP8PerChannel
+import vllm_gaudi.extension.ops as hpu_ops
+from vllm_gaudi.extension.ops import VllmMixtureOfExpertsOpFP8PerChannel
 
 SUPPORTED_STRATEGIES = [
     QuantizationStrategy.CHANNEL, QuantizationStrategy.TENSOR

vllm_hpu/ops/hpu_fp8.py → vllm_gaudi/ops/hpu_fp8.py

@@ -1,7 +1,7 @@
 from typing import Callable, Optional
 
 import torch
-from vllm_hpu import envs
+from vllm_gaudi import envs
 from vllm.model_executor.custom_op import CustomOp
 from vllm.model_executor.layers.fused_moe.layer import FusedMoE
 
@@ -10,9 +10,9 @@
                                                          OrigFp8LinearMethod,
                                                          Fp8MoEMethod,
                                                          Fp8Config)
-import vllm_hpu.extension.ops as hpu_ops
-from vllm_hpu.extension.ops import (VllmMixtureOfExpertsOpFP8PerChannel,
-                                    VllmMixtureOfExpertsOpFP8)
+import vllm_gaudi.extension.ops as hpu_ops
+from vllm_gaudi.extension.ops import (VllmMixtureOfExpertsOpFP8PerChannel,
+                                      VllmMixtureOfExpertsOpFP8)
 
 
 class Fp8LinearMethod(OrigFp8LinearMethod):

vllm_hpu/ops/hpu_fused_moe.py → vllm_gaudi/ops/hpu_fused_moe.py

@@ -3,7 +3,7 @@
 import torch
 from vllm.model_executor.layers.fused_moe.layer import (
     FusedMoE, UnquantizedFusedMoEMethod)
-from vllm_hpu.extension.ops import (VllmMixtureOfExpertsOp)
+from vllm_gaudi.extension.ops import (VllmMixtureOfExpertsOp)
 
 
 @UnquantizedFusedMoEMethod.register_oot

vllm_hpu/ops/hpu_layernorm.py → vllm_gaudi/ops/hpu_layernorm.py

@@ -12,7 +12,7 @@ def forward_oot(
         x: torch.Tensor,
         residual: Optional[torch.Tensor] = None,
     ) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
-        from vllm_hpu.extension.kernels import rms_norm
+        from vllm_gaudi.extension.kernels import rms_norm
         HPUFusedRMSNorm = rms_norm()
         if residual is not None:
             orig_shape = x.shape