[Deps] Bump FlashInfer to 0.7.0.post1 (#40709)
Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai>
This commit is contained in:
co-authored by
Mohammad Angkad
parent
4dfc0ac131
commit
eae808903f
@@ -39,7 +39,7 @@ dependencies = [
|
||||
"einops",
|
||||
"fastapi",
|
||||
"flash-attn-4>=4.0.0b18",
|
||||
"flashinfer_python[cu13]==0.6.18", # keep it aligned with jit-cache version in Dockerfile
|
||||
"flashinfer_python[cu13]==0.7.0.post1", # keep it aligned with jit-cache version in Dockerfile
|
||||
"gguf",
|
||||
"humming-kernels[cu13]==0.1.12",
|
||||
"interegular",
|
||||
@@ -52,7 +52,7 @@ dependencies = [
|
||||
"ninja",
|
||||
"numba==0.65.1",
|
||||
"numpy",
|
||||
"nvidia-cutlass-dsl[cu13]==4.6.2",
|
||||
"nvidia-cutlass-dsl[cu13]==4.8.0",
|
||||
"nvidia-mathdx==25.6.0",
|
||||
"nvidia-ml-py",
|
||||
"nvshmem4py-cu13",
|
||||
@@ -70,7 +70,7 @@ dependencies = [
|
||||
"pydantic",
|
||||
"python-multipart",
|
||||
"pyzmq>=25.1.2",
|
||||
"quack-kernels==0.6.4",
|
||||
"quack-kernels==0.6.5",
|
||||
"regex",
|
||||
"requests",
|
||||
"scipy",
|
||||
|
||||
@@ -22,17 +22,17 @@ classifiers = [
|
||||
]
|
||||
|
||||
dependencies = [
|
||||
"nvidia-cutlass-dsl==4.6.2",
|
||||
"nvidia-cutlass-dsl[cu13]==4.8.0",
|
||||
"torch",
|
||||
"einops",
|
||||
"typing_extensions",
|
||||
"apache-tvm-ffi>=0.1.5,<0.2",
|
||||
"torch-c-dlpack-ext",
|
||||
"quack-kernels==0.6.4",
|
||||
"quack-kernels==0.6.5",
|
||||
]
|
||||
|
||||
[project.optional-dependencies]
|
||||
cu13 = ["nvidia-cutlass-dsl[cu13]==4.6.2"]
|
||||
cu13 = ["nvidia-cutlass-dsl[cu13]==4.8.0"]
|
||||
dev = [
|
||||
"pytest",
|
||||
"pytest-xdist",
|
||||
|
||||
@@ -9,7 +9,6 @@ from typing import Callable, NamedTuple, Optional, Tuple, Type, overload
|
||||
|
||||
import cutlass
|
||||
import cutlass.cute as cute
|
||||
import quack.activation
|
||||
from cutlass import Float32, Int32, const_expr
|
||||
from cutlass._mlir.dialects import llvm, nvvm
|
||||
from cutlass.cute import FastDivmodDivisor
|
||||
@@ -925,10 +924,10 @@ def ex2_emulation_2(
|
||||
)
|
||||
# The integer floor of x & y are now in the last 8 bits of xy_rounded
|
||||
# We want the next 2 ops to round to nearest even. The rounding mode is important.
|
||||
xy_rounded_back = quack.activation.sub_packed_f32x2(
|
||||
xy_rounded_back = cute.arch.sub_packed_f32x2(
|
||||
xy_rounded, (fp32_round_int, fp32_round_int)
|
||||
)
|
||||
xy_frac = quack.activation.sub_packed_f32x2(xy_clamped, xy_rounded_back)
|
||||
xy_frac = cute.arch.sub_packed_f32x2(xy_clamped, xy_rounded_back)
|
||||
xy_frac_ex2 = evaluate_polynomial_2(*xy_frac, POLY_EX2[poly_degree], loc=loc, ip=ip)
|
||||
x_out = combine_int_frac_ex2(xy_rounded[0], xy_frac_ex2[0], loc=loc, ip=ip)
|
||||
y_out = combine_int_frac_ex2(xy_rounded[1], xy_frac_ex2[1], loc=loc, ip=ip)
|
||||
|
||||
+1
-1
@@ -109,7 +109,7 @@ pip install git+https://github.com/thu-ml/SpargeAttn.git --no-build-isolation
|
||||
SM120 requires FlashInfer's CuTe-DSL SM120 Sage backend and
|
||||
`quantize_sage_qkv_sm120`, added in
|
||||
[FlashInfer #4691](https://github.com/flashinfer-ai/flashinfer/pull/4691).
|
||||
The default FlashInfer `0.6.18` pin does not include these APIs. After installing
|
||||
The default FlashInfer `0.7.0.post1` pin does not include these APIs. After installing
|
||||
SGLang, install the FlashInfer source revision used for SM120 validation:
|
||||
|
||||
```bash
|
||||
|
||||
@@ -1848,7 +1848,7 @@ def _set_envs_and_config(server_args: ServerArgs):
|
||||
):
|
||||
assert_pkg_version(
|
||||
"flashinfer_python",
|
||||
"0.6.18",
|
||||
"0.7.0.post1",
|
||||
"Please uninstall the old version and "
|
||||
"reinstall the latest version by following the instructions "
|
||||
"at https://docs.flashinfer.ai/installation.html.",
|
||||
|
||||
@@ -934,6 +934,11 @@ class FlashInferMLAIndicesUpdaterDecode:
|
||||
valid.copy_(translator.translate_dcp_read_ids(valid))
|
||||
else:
|
||||
kv_indptr, kv_indices = spec_info.kv_indptr, spec_info.kv_indices
|
||||
if not init_metadata_replay:
|
||||
# FlashInfer >= 0.7.0 rejects a plan whose kv_len_arr disagrees with
|
||||
# kv_indptr, and a draft branch spans its prefix plus one per step.
|
||||
# Replay plans through fast_mla_decode_plan, which skips that check.
|
||||
kv_lens = (kv_indptr[1:] - kv_indptr[:-1]).to(torch.int32)
|
||||
|
||||
if not init_metadata_replay:
|
||||
wrapper.plan(
|
||||
|
||||
@@ -2264,16 +2264,7 @@ def assert_pkg_version(pkg: str, min_version: str, message: str):
|
||||
|
||||
|
||||
def check_pkg_version_at_least(pkg: str, min_version: str) -> bool:
|
||||
"""
|
||||
Check if a package is installed and meets the minimum version requirement.
|
||||
|
||||
Args:
|
||||
pkg: Package name (distribution name, e.g., "flashinfer-python")
|
||||
min_version: Minimum version required (e.g., "0.6.18")
|
||||
|
||||
Returns:
|
||||
True if package is installed and version >= min_version, False otherwise
|
||||
"""
|
||||
"""Check if a package is installed and meets the minimum version requirement."""
|
||||
if _should_skip_kernel_pkg_version_check(pkg):
|
||||
return True
|
||||
|
||||
|
||||
Reference in New Issue
Block a user