MCPcopy Index your code
hub / github.com/FlashSampling/FlashSampling / make_image

Function make_image

src/fused_mm_sampling/modal_lib/utils.py:35–58  ·  view source on GitHub ↗
()

Source from the content-addressed store, hash-verified

33
34
35def make_image():
36 # Pytorch 2.11.0 ships triton 3.6.0, whose bundled `ptxas-blackwell`
37 # supports sm_103a (B300). We install deps via `uv pip install --system`
38 # (mirroring `modal_vllm_benchmark.py`) so uv preserves the base image's
39 # pre-installed torch 2.11.0+cu130 and triton 3.6.0 instead of re-resolving
40 # them from PyPI (which yields torch 2.9.1+cu128 + triton 3.5.1 without
41 # ptxas-blackwell).
42 deps: list[str] = [
43 "flashinfer-python",
44 "pandas",
45 "pydantic-settings",
46 "matplotlib",
47 "nvtx",
48 "llnl-hatchet",
49 "scipy",
50 "'cuda-bench[cu13]'",
51 "cupti-python",
52 ]
53 deps_str: str = " ".join(deps)
54 return (
55 modal.Image.from_registry("pytorch/pytorch:2.11.0-cuda13.0-cudnn9-devel")
56 .run_commands("pip install --break-system-packages uv")
57 .run_commands(f"uv pip install --system {deps_str}")
58 )
59
60
61VLLM_FORK_BRANCH = "feature/fmms-sampler"

Callers 4

modal_nsys_test.pyFile · 0.85
modal_ncu.pyFile · 0.85
modal_nsys.pyFile · 0.85
modal_example.pyFile · 0.85

Calls

no outgoing calls

Tested by

no test coverage detected