Uploaded using `kernel-builder`.

Browse files

Files changed (11) hide show

benchmarks/benchmark.py +90 -0
build/torch211-cxx11-cu130-x86_64-linux/__init__.py +101 -0
build/torch211-cxx11-cu130-x86_64-linux/_audio_codebook_primitives_cuda_bcb0782.abi3.so +3 -0
build/torch211-cxx11-cu130-x86_64-linux/_ops.py +9 -0
build/torch211-cxx11-cu130-x86_64-linux/audio_codebook_primitives/__init__.py +26 -0
build/torch211-cxx11-cu130-x86_64-linux/metadata.json +20 -0
build/torch212-cxx11-cu130-x86_64-linux/__init__.py +101 -0
build/torch212-cxx11-cu130-x86_64-linux/_audio_codebook_primitives_cuda_bcb0782.abi3.so +3 -0
build/torch212-cxx11-cu130-x86_64-linux/_ops.py +9 -0
build/torch212-cxx11-cu130-x86_64-linux/audio_codebook_primitives/__init__.py +26 -0
build/torch212-cxx11-cu130-x86_64-linux/metadata.json +20 -0

benchmarks/benchmark.py ADDED Viewed

	@@ -0,0 +1,90 @@

+#!/usr/bin/env python3
+"""Higgs delayed-codebook benchmark against eager/compile and raw op."""
+from __future__ import annotations
+import argparse
+import importlib
+import sys
+from pathlib import Path
+import torch
+def elapsed_us(fn, warmup, iterations):
+    for _ in range(warmup):
+        fn()
+    torch.cuda.synchronize()
+    start = torch.cuda.Event(enable_timing=True)
+    end = torch.cuda.Event(enable_timing=True)
+    start.record()
+    for _ in range(iterations):
+        fn()
+    end.record()
+    end.synchronize()
+    return start.elapsed_time(end) * 1000.0 / iterations
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--backend", choices=("source", "installed"), default="source")
+    parser.add_argument("--artifact")
+    parser.add_argument("--warmup", type=int, default=50)
+    parser.add_argument("--iterations", type=int, default=500)
+    args = parser.parse_args()
+    if args.backend == "source":
+        tests = Path(__file__).resolve().parents[1] / "tests"
+        sys.path.insert(0, str(tests))
+        from test_audio_codebook_primitives import load_source_ops
+        ops = load_source_ops()
+    else:
+        if args.artifact:
+            sys.path.insert(0, args.artifact)
+        ops = importlib.import_module("audio_codebook_primitives")
+    c, v, h, delay, boc = 8, 1026, 1024, 7, 1024
+    logits = torch.randn((c, v), device="cuda", dtype=torch.bfloat16)
+    codebook = torch.randn((c, v, h), device="cuda", dtype=torch.bfloat16)
+    index = torch.arange(c, device="cuda")
+    active = index <= delay
+    boc_tensor = torch.full((c,), boc, device="cuda", dtype=torch.int64)
+    def eager():
+        codes = torch.where(active, logits.argmax(dim=1), boc_tensor)
+        embedding = codebook[index, codes].float().sum(dim=0).bfloat16()
+        return codes, embedding
+    compiled = torch.compile(eager, fullgraph=True)
+    codes = torch.empty(c, device="cuda", dtype=torch.int64)
+    embedding = torch.empty(h, device="cuda", dtype=torch.bfloat16)
+    def wrapper():
+        return ops.delayed_codebook_argmax_embed_bf16(
+            logits, codebook, delay=delay, boc=boc,
+            codes=codes, embedding=embedding
+        )
+    namespace = ops.ops
+    def raw():
+        namespace.delayed_codebook_argmax_embed_bf16(
+            logits, codebook, delay, boc, codes, embedding
+        )
+    expected = eager()
+    actual = wrapper()
+    torch.testing.assert_close(actual[0], expected[0], rtol=0, atol=0)
+    torch.testing.assert_close(actual[1], expected[1], rtol=0, atol=0)
+    rows = {
+        "torch_eager_us": elapsed_us(eager, args.warmup, args.iterations),
+        "torch_compile_us": elapsed_us(compiled, args.warmup, args.iterations),
+        "hub_wrapper_us": elapsed_us(wrapper, args.warmup, args.iterations),
+        "raw_native_op_us": elapsed_us(raw, args.warmup, args.iterations),
+    }
+    for name, value in rows.items():
+        print(f"{name}={value:.3f}")
+if __name__ == "__main__":
+    main()

build/torch211-cxx11-cu130-x86_64-linux/__init__.py ADDED Viewed

	@@ -0,0 +1,101 @@

+"""FlashRT delayed-codebook selection and embedding kernels."""
+from __future__ import annotations
+from typing import Optional
+import torch
+from ._ops import add_op_namespace_prefix, ops
+def _check(logits, codebook, delay: int, boc: int, codes, embedding) -> None:
+    if logits.dim() != 2:
+        raise RuntimeError("logits must have shape (num_codebooks, codebook_vocab)")
+    if codebook.dim() != 3 or codebook.shape[:2] != logits.shape:
+        raise RuntimeError("codebook must have shape (num_codebooks, codebook_vocab, hidden)")
+    if codes.shape != (logits.shape[0],):
+        raise RuntimeError("codes must have shape (num_codebooks,)")
+    if embedding.shape != (codebook.shape[2],):
+        raise RuntimeError("embedding must have shape (hidden,)")
+    if not 0 <= delay <= logits.shape[0]:
+        raise RuntimeError("delay must be in [0, num_codebooks]")
+    if not 0 <= boc < logits.shape[1]:
+        raise RuntimeError("boc must be a valid codebook index")
+@torch.library.register_fake(
+    add_op_namespace_prefix("delayed_codebook_argmax_embed_bf16")
+)
+def _argmax_fake(logits, codebook, delay: int, boc: int, codes, embedding) -> None:
+    _check(logits, codebook, delay, boc, codes, embedding)
+    return None
+@torch.library.register_fake(
+    add_op_namespace_prefix("delayed_codebook_sample_embed_bf16")
+)
+def _sample_fake(
+    logits, codebook, delay: int, boc: int, temperature: float,
+    seed: int, step: int, codes, embedding
+) -> None:
+    _check(logits, codebook, delay, boc, codes, embedding)
+    if temperature <= 0:
+        raise RuntimeError("temperature must be strictly positive")
+    if step < 0:
+        raise RuntimeError("step must be non-negative")
+    return None
+def _outputs(logits, codebook, codes, embedding):
+    if codes is None:
+        codes = torch.empty(
+            (logits.shape[0],), device=logits.device, dtype=torch.int64
+        )
+    if embedding is None:
+        embedding = torch.empty(
+            (codebook.shape[2],), device=codebook.device, dtype=torch.bfloat16
+        )
+    return codes, embedding
+def delayed_codebook_argmax_embed_bf16(
+    logits: torch.Tensor,
+    codebook: torch.Tensor,
+    *,
+    delay: int,
+    boc: int,
+    codes: Optional[torch.Tensor] = None,
+    embedding: Optional[torch.Tensor] = None,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    codes, embedding = _outputs(logits, codebook, codes, embedding)
+    ops.delayed_codebook_argmax_embed_bf16(
+        logits, codebook, int(delay), int(boc), codes, embedding
+    )
+    return codes, embedding
+def delayed_codebook_sample_embed_bf16(
+    logits: torch.Tensor,
+    codebook: torch.Tensor,
+    *,
+    delay: int,
+    boc: int,
+    temperature: float,
+    seed: int,
+    step: int,
+    codes: Optional[torch.Tensor] = None,
+    embedding: Optional[torch.Tensor] = None,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    codes, embedding = _outputs(logits, codebook, codes, embedding)
+    ops.delayed_codebook_sample_embed_bf16(
+        logits, codebook, int(delay), int(boc), float(temperature),
+        int(seed), int(step), codes, embedding
+    )
+    return codes, embedding
+__all__ = [
+    "delayed_codebook_argmax_embed_bf16",
+    "delayed_codebook_sample_embed_bf16",
+]

build/torch211-cxx11-cu130-x86_64-linux/_audio_codebook_primitives_cuda_bcb0782.abi3.so ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:6ad6d97d9bda4b81bf8e78f10167f7de84cfed2c25045dda31fe95503f57aae3
+size 315320

build/torch211-cxx11-cu130-x86_64-linux/_ops.py ADDED Viewed

	@@ -0,0 +1,9 @@

+import torch
+from . import _audio_codebook_primitives_cuda_bcb0782
+ops = torch.ops._audio_codebook_primitives_cuda_bcb0782
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_audio_codebook_primitives_cuda_bcb0782::{op_name}"

build/torch211-cxx11-cu130-x86_64-linux/audio_codebook_primitives/__init__.py ADDED Viewed

	@@ -0,0 +1,26 @@

+import ctypes
+import importlib.util
+import sys
+from pathlib import Path
+from types import ModuleType
+def _import_from_path(file_path: Path) -> ModuleType:
+    # We cannot use the module name as-is, after adding it to `sys.modules`,
+    # it would also be used for other imports. So, we make a module name that
+    # depends on the path for it to be unique using the hex-encoded hash of
+    # the path.
+    path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
+    module_name = path_hash
+    spec = importlib.util.spec_from_file_location(module_name, file_path)
+    if spec is None:
+        raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
+    module = importlib.util.module_from_spec(spec)
+    if module is None:
+        raise ImportError(f"Cannot load module {module_name} from spec")
+    sys.modules[module_name] = module
+    spec.loader.exec_module(module)  # type: ignore
+    return module
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))

build/torch211-cxx11-cu130-x86_64-linux/metadata.json ADDED Viewed

	@@ -0,0 +1,20 @@

+{
+  "name": "audio-codebook-primitives",
+  "id": "_audio_codebook_primitives_cuda_bcb0782",
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "10.0",
+      "11.0",
+      "12.0",
+      "12.1",
+      "8.0",
+      "8.6",
+      "8.9",
+      "9.0"
+    ]
+  }
+}

build/torch212-cxx11-cu130-x86_64-linux/__init__.py ADDED Viewed

	@@ -0,0 +1,101 @@

+"""FlashRT delayed-codebook selection and embedding kernels."""
+from __future__ import annotations
+from typing import Optional
+import torch
+from ._ops import add_op_namespace_prefix, ops
+def _check(logits, codebook, delay: int, boc: int, codes, embedding) -> None:
+    if logits.dim() != 2:
+        raise RuntimeError("logits must have shape (num_codebooks, codebook_vocab)")
+    if codebook.dim() != 3 or codebook.shape[:2] != logits.shape:
+        raise RuntimeError("codebook must have shape (num_codebooks, codebook_vocab, hidden)")
+    if codes.shape != (logits.shape[0],):
+        raise RuntimeError("codes must have shape (num_codebooks,)")
+    if embedding.shape != (codebook.shape[2],):
+        raise RuntimeError("embedding must have shape (hidden,)")
+    if not 0 <= delay <= logits.shape[0]:
+        raise RuntimeError("delay must be in [0, num_codebooks]")
+    if not 0 <= boc < logits.shape[1]:
+        raise RuntimeError("boc must be a valid codebook index")
+@torch.library.register_fake(
+    add_op_namespace_prefix("delayed_codebook_argmax_embed_bf16")
+)
+def _argmax_fake(logits, codebook, delay: int, boc: int, codes, embedding) -> None:
+    _check(logits, codebook, delay, boc, codes, embedding)
+    return None
+@torch.library.register_fake(
+    add_op_namespace_prefix("delayed_codebook_sample_embed_bf16")
+)
+def _sample_fake(
+    logits, codebook, delay: int, boc: int, temperature: float,
+    seed: int, step: int, codes, embedding
+) -> None:
+    _check(logits, codebook, delay, boc, codes, embedding)
+    if temperature <= 0:
+        raise RuntimeError("temperature must be strictly positive")
+    if step < 0:
+        raise RuntimeError("step must be non-negative")
+    return None
+def _outputs(logits, codebook, codes, embedding):
+    if codes is None:
+        codes = torch.empty(
+            (logits.shape[0],), device=logits.device, dtype=torch.int64
+        )
+    if embedding is None:
+        embedding = torch.empty(
+            (codebook.shape[2],), device=codebook.device, dtype=torch.bfloat16
+        )
+    return codes, embedding
+def delayed_codebook_argmax_embed_bf16(
+    logits: torch.Tensor,
+    codebook: torch.Tensor,
+    *,
+    delay: int,
+    boc: int,
+    codes: Optional[torch.Tensor] = None,
+    embedding: Optional[torch.Tensor] = None,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    codes, embedding = _outputs(logits, codebook, codes, embedding)
+    ops.delayed_codebook_argmax_embed_bf16(
+        logits, codebook, int(delay), int(boc), codes, embedding
+    )
+    return codes, embedding
+def delayed_codebook_sample_embed_bf16(
+    logits: torch.Tensor,
+    codebook: torch.Tensor,
+    *,
+    delay: int,
+    boc: int,
+    temperature: float,
+    seed: int,
+    step: int,
+    codes: Optional[torch.Tensor] = None,
+    embedding: Optional[torch.Tensor] = None,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    codes, embedding = _outputs(logits, codebook, codes, embedding)
+    ops.delayed_codebook_sample_embed_bf16(
+        logits, codebook, int(delay), int(boc), float(temperature),
+        int(seed), int(step), codes, embedding
+    )
+    return codes, embedding
+__all__ = [
+    "delayed_codebook_argmax_embed_bf16",
+    "delayed_codebook_sample_embed_bf16",
+]

build/torch212-cxx11-cu130-x86_64-linux/_audio_codebook_primitives_cuda_bcb0782.abi3.so ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:5d2365e676c2962d167dccbf1d71fc1e0da4a243cafc36d866f12362d1af2020
+size 321408

build/torch212-cxx11-cu130-x86_64-linux/_ops.py ADDED Viewed

	@@ -0,0 +1,9 @@

+import torch
+from . import _audio_codebook_primitives_cuda_bcb0782
+ops = torch.ops._audio_codebook_primitives_cuda_bcb0782
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_audio_codebook_primitives_cuda_bcb0782::{op_name}"

build/torch212-cxx11-cu130-x86_64-linux/audio_codebook_primitives/__init__.py ADDED Viewed

	@@ -0,0 +1,26 @@

+import ctypes
+import importlib.util
+import sys
+from pathlib import Path
+from types import ModuleType
+def _import_from_path(file_path: Path) -> ModuleType:
+    # We cannot use the module name as-is, after adding it to `sys.modules`,
+    # it would also be used for other imports. So, we make a module name that
+    # depends on the path for it to be unique using the hex-encoded hash of
+    # the path.
+    path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
+    module_name = path_hash
+    spec = importlib.util.spec_from_file_location(module_name, file_path)
+    if spec is None:
+        raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
+    module = importlib.util.module_from_spec(spec)
+    if module is None:
+        raise ImportError(f"Cannot load module {module_name} from spec")
+    sys.modules[module_name] = module
+    spec.loader.exec_module(module)  # type: ignore
+    return module
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))

build/torch212-cxx11-cu130-x86_64-linux/metadata.json ADDED Viewed

	@@ -0,0 +1,20 @@

+{
+  "name": "audio-codebook-primitives",
+  "id": "_audio_codebook_primitives_cuda_bcb0782",
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "10.0",
+      "11.0",
+      "12.0",
+      "12.1",
+      "8.0",
+      "8.6",
+      "8.9",
+      "9.0"
+    ]
+  }
+}