Skip to content

Sley 2.1.7

Choose a tag to compare

@GreyforgeLabs GreyforgeLabs released this 28 Oct 03:15
· 133 commits to main since this release
Immutable release. Only release title and notes can be modified.

Sley 2.1.7 is a performance point release. Python-dialect list indexing by a loop counter no longer converts the index on every access (#69). The VM word tier now reads vectors (len, indexing) and fuses more of each burst (#70). On the shipped musl binary, seven of the nine programs that also ran on 2.1.6 are 1.09-1.26x faster, and the list-indexing inversion count now finishes where 2.1.6 ran out of fuel. With process start included, Sley is faster than CPython 3.13 on ten of eleven benchmarks and slower on digitsum (1.10x). fib(30) is about 4% slower than in 2.1.6. C, Rust, Go, LuaJIT and Lua 5.4 remain faster on every program. Fuel and instruction counts match call --reference. Over 14 tasks of a scripted write-test-fix loop, the Python dialect uses 0.64x the tokens of Python, the fewest of the six languages measured.

Release notes · Benchmarks · Token cost · Quickstart

Changes

  • Loop-counter index shadow (#69): a range loop whose body indexes a list by its int counter carries a u64 copy of the counter, including i ± k offsets. An enumerate loop gets an i64 copy of its index. Out-of-range and negative indexes fail as before.
  • Word-tier vector reads and fusion (#70): Vec<word> and Option<word> values are now word-tier types, with new VecLen/VecGet ops and Some/None switches on the Ok/Err edges. Conditionals fuse one side whenever either side can fuse, guards and edge moves merge into their producer steps, and moves into slots that are never read are dropped.

Benchmarks (wall ms, hyperfine median [min-max], release-archive musl binaries, one session)

benchmark Sley 2.1.6 Sley 2.1.7 2.1.6 / 2.1.7 CPython 3.13.5 PyPy 7.3.19 Lua 5.4.7 LuaJIT 2.1 Node 20.19 Ruby 3.3.8 Go 1.27 C gcc -O2 Rust -O
fib 30 73.2 [72.6-105] 76.4 [75.2-85.5] 0.96x 82.7 [81.2-85.0] 169 [164-186] 37.2 [36.8-45.2] 7.8 [7.4-19.7] 73.2 [70.4-108] 93.9 [93.2-148] 4.2 [3.9-57.4] 1.4 [1.3-3.5] 2.4 [2.3-76.1]
loopsum 1e6 40.7 [40.2-129] 34.4 [33.8-44.2] 1.18x 57.8 [55.7-65.5] 30.9 [30.5-34.0] 8.4 [8.0-9.0] 8.6 [8.4-9.1] 78.8 [76.6-83.0] 69.9 [68.9-75.4] 3.9 [3.6-13.6] 3.4 [3.3-28.3] 3.7 [3.5-84.7]
collatz 5000 30.8 [30.3-68.0] 27.1 [26.7-40.3] 1.14x 35.3 [34.8-41.8] 29.3 [28.7-34.6] 8.0 [7.7-11.2] 2.5 [2.3-2.9] 67.7 [65.3-70.6] 55.9 [54.3-72.1] 1.2 [1.2-12.3] 1.1 [1.1-91.6] 0.81 [0.77-1.7]
primes 30000 30.4 [30.0-49.7] 27.9 [27.5-36.5] 1.09x 40.9 [40.4-42.9] 31.1 [30.5-38.1] 7.6 [7.4-8.4] 2.5 [2.4-3.0] 67.7 [65.3-70.1] 58.5 [56.4-68.5] 2.1 [2.0-4.1] 1.7 [1.7-27.6] 1.3 [1.2-7.9]
gcdsum 400 31.2 [30.8-86.2] 25.6 [25.2-30.4] 1.22x 42.2 [41.8-43.1] 69.2 [67.0-83.7] 12.1 [11.9-12.9] 6.1 [5.9-7.4] 72.5 [70.6-83.8] 63.7 [62.5-78.2] 4.6 [4.3-111] 4.0 [3.9-8.9] 3.4 [3.4-15.9]
tak 24 16 8 68.9 [68.0-199] 69.6 [67.8-82.3] 0.99x 80.0 [78.4-86.5] 249 [238-275] 35.9 [35.0-52.2] 8.7 [7.8-11.7] 70.3 [68.9-81.2] 94.2 [93.7-260] 5.2 [5.0-5.6] 3.0 [3.0-10.1] 3.4 [3.3-9.8]
triples 120 15.4 [14.9-70.8] 12.2 [11.8-14.5] 1.26x 25.5 [25.0-30.9] 29.1 [28.7-32.4] 4.7 [4.4-16.1] 1.2 [0.92-9.9] 68.3 [65.8-87.3] 50.6 [49.9-54.0] 0.72 [0.68-1.7] 0.46 [0.43-1.2] 0.52 [0.49-1.1]
digitsum 150000 59.8 [59.1-82.6] 53.1 [52.5-55.8] 1.12x 48.5 [47.0-56.6] 30.9 [29.8-191] 9.5 [9.1-13.3] 2.3 [2.2-3.6] 72.9 [68.9-165] 66.2 [64.5-80.1] 1.2 [1.1-2.8] 2.3 [2.2-2.9] 0.91 [0.87-2.2]
isqrt 50000 47.3 [46.6-49.2] 39.9 [39.5-42.0] 1.19x 57.7 [56.9-59.0] 33.5 [32.7-85.6] 13.4 [13.3-84.7] 3.7 [3.6-4.1] 70.6 [69.1-77.9] 68.1 [67.5-131] 1.2 [1.2-61.3] 2.2 [2.1-55.7] 0.98 [0.94-53.7]
inversions 1000 98.9 [96.6-110] (out of fuel) 19.7 [19.4-57.0] 2.1.6 out of fuel 30.0 [29.5-35.9] 30.1 [29.9-34.4] 5.8 [5.7-6.9] 2.5 [2.4-3.2] 67.9 [65.6-75.9] 55.9 [54.7-59.1] 0.85 [0.79-27.0] 1.2 [1.1-44.8] 0.63 [0.60-79.6]
inversions 400 97.3 [96.2-121] (out of fuel) 8.5 [8.4-12.3] 2.1.6 out of fuel 16.9 [16.6-79.4] 30.0 [29.4-91.2] 1.6 [1.5-56.0] 1.1 [1.0-12.5] 69.7 [66.1-91.3] 43.1 [42.6-56.7] 0.66 [0.62-1.4] 0.44 [0.41-4.8] 0.47 [0.43-8.4]

These are development measurements on one 8-vCPU cloud VM, pinned to one vCPU, with 2 warmups and at least 10 runs. They are not a GA claim.

  • Regression: fib(30) is about 4% slower than in 2.1.6 (76.4 vs 73.2 ms). tak, the other call-heavy program, is unchanged.
  • Where Sley is slower: CPython is faster on digitsum, and PyPy on loopsum, digitsum and isqrt. Node is slightly faster on fib(30) (1.04x). C, Rust, Go, LuaJIT and Lua 5.4 are 1.9-58x faster everywhere.
  • Limits: sley-agent call has a 100M fuel ceiling, which keeps the inputs small.

Token cost of a write-test-fix loop (o200k_base, 14 tasks, scripted)

scenario Python Sley 2.1.7 dialect C Rust Go TypeScript
clean 2054 1953 (0.95x) 2576 (1.25x) 3362 (1.64x) 3315 (1.61x) 2490 (1.21x)
logic bug 4063 3093 (0.76x) 3475 (0.86x) 5581 (1.37x) 4460 (1.10x) 5703 (1.40x)
typo 6064 2768 (0.46x) 4412 (0.73x) 5220 (0.86x) 3833 (0.63x) 3065 (0.51x)
all three 12181 7814 (0.64x) 10463 (0.86x) 14163 (1.16x) 11608 (0.95x) 11258 (0.92x)

These are scripted tokenizer counts with no model in the loop, not model spend. The one-time 407-token help python guide is excluded. C's assert stops at the first failure, and TypeScript uses Node's quiet dot reporter.

The kernel, SSMC1, SCB1, SMP1 rules, the instruction set, admission and commit semantics are unchanged. No third-party dependency was added. The native test supervisor is still source only, so committing a candidate that selects TestCases still refuses with TXN_TEST_EVIDENCE_UNSUPPORTED.

Two maintainer checks, check_candidate_result_contract.py and check_transaction_contract.py, were waived for this release. They need the Sley 2.0 design specification, which is unavailable off the maintainer host. The waiver is carried forward from the 2.1.2 operator decision, and the operator has not re-confirmed it. This is not a pass. The persistent-fuzz proofs use the documented toolchain override (Debian clang 18.1.8).

Linux x86_64

sley-2.1.7-linux-x86_64.tar.gz contains static musl sley and sley-agent binaries. It was built from f5ef71254c57cb8fdbef33bc6077287cc1cdaa43, and two clean builds on the primary host produced identical archives.

  • Size: 7,677,962 bytes; 17 archive members
  • SHA-256: a44b76e070228e2924c7efe16319d5339b233abf1e45986431c59a4c88460923
sha256sum -c SHA256SUMS
tar xzf sley-2.1.7-linux-x86_64.tar.gz
cd sley-2.1.7-linux-x86_64
./bin/sley version
./bin/sley-agent help python
python3 demo/run_demo.py

ga_claimed remains false. Cross-host reproducibility is not claimed, and provenance is unsigned.