1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
// SPDX-License-Identifier: MIT
//! Phase-4 acceptance: real bit-exact replay against vLLM ≥ 0.10 in
//! batch-invariant mode, on a CUDA host.
//!
//! Gated behind `$PF_HAS_GPU=1`. The test boots vLLM as a subprocess with
//! `--enforce-deterministic`, runs Llama-3-8B for ~50 tokens, snapshots,
//! kills the worker, restores into a fresh worker, runs another 100 tokens
//! on each branch, and asserts logit-bit-identical output.
//!
//! On hosts without `PF_HAS_GPU=1` (the build-host default) the test
//! `eprintln!`-skips so the workspace gate stays green. Operators on a
//! Hopper-class box can run this without code changes.