vh/esh-pfi-infrastructure
1
0
Fork 0
Code Issues 7 Pull Requests Actions Packages Projects Releases Wiki Activity

feat(semif): 0.1.3 — order averaging, fast kernels, bug-hunt hardening (Prime)

Browse Source 
Order averaging (Prime, after the 739aa03 spike):
- A decision may set orderings: rotations|all (all only for <= 4 options). Every
  ordering goes to the engine in one shared batch.
- The reply keeps each native result and adds combined {probabilities (log-mean),
  top, agreement, spread}.
- Through the service on SemIf's labelled sets (252 rows): 78.6% -> 88.1%
  (group-bootstrap 95% CI +5.1..+14.3). Unanimous agreement is 94.5% accurate.

Fast kernels: flash-linear-attention 0.5.2 and causal-conv1d 1.7.0 are now the
default build. A/B on the empty GPU 3:
- parity with upstream went from 142/144 to 144/144;
- a ~2k-token /decide went from 169 to 92 ms server-side;
- short 3-rotation batches cost ~3-6 ms more.
triton builds a C shim at runtime, so the image carries gcc. Without it the
warm-up failed and startup failed closed.

Heid bug-hunt panel (4/4 arms, thread 01M3H3F4RR7XBP90KQ3A39H4SX), folded:
- Startup validation: VRAM cap 0 no longer means uncapped (C1); limits must be
  >= 1 (S1); the token must be visible ASCII (S2); the calibration file must
  exist and parse, with T in [0.05, 20] (S8, and C3's NaN leg).
- The body limit is checked before a chunk is kept, and a Unicode-digit
  Content-Length no longer crashes (C2, S3).
- Failures while building the response now get the 500 envelope (C3).
- 429 busy past SEMIF_MAX_QUEUE requests in progress (C6).
- The engine releases memory on every non-validation failure, unchained after
  gc; an empty OOM message is handled; 'out of memory' RuntimeErrors map to 503
  (C4, C5, S9).
- The entry point forces HF_HUB_OFFLINE (S10). README wording fixed (S5, S6).
- New guard tests close the gaps the arms' mutation grids exposed: early stop of
  the body read, a shared-route lock, calibration pass-through, the gc cycle,
  the exact caps, TorchEngine.load's arch and device checks, and the offline
  entry point.
86 tests.

Deployed on fv-ml1 GPU 1: parity 144/144, OOM and burst release verified, shared
capacity 63/51/26/16 rows at ~140/520/1960/3900 prefix tokens.
This commit is contained in:
vh
2026-09-27 03:27:15 -07:00
parent d7ad235365
commit 77b8cb449c
22 changed files with 1018 additions and 92 deletions
Download Patch File Download Diff File Expand all files Collapse all files
services/semif-serve/uv.lock
Generated
+80 -2
View File
@@ -51,6 +51,26 @@ wheels = [
{ url = "https://files.pythonhosted.org/packages/12/b8/4bd346e22b28902df4d651910f5242c28d84e4a5c2435ca5c3f797ed7e2e/anyio-4.15.1-py3-none-any.whl", hash = "sha256:6152fdbbf9a77fdec97731721bebf7c4c44f7c29b424b0065826173efc7ed101", size = 132079 },
]
[[package]]
name = "causal-conv1d"
version = "1.7.0"
source = { url = "https://github.com/Dao-AILab/causal-conv1d/releases/download/v1.7.0/causal_conv1d-1.7.0+cu12torch2.10cxx11abiTRUE-cp312-cp312-linux_x86_64.whl" }
dependencies = [
{ name = "ninja" },
{ name = "packaging" },
{ name = "torch" },
]
wheels = [
{ url = "https://github.com/Dao-AILab/causal-conv1d/releases/download/v1.7.0/causal_conv1d-1.7.0+cu12torch2.10cxx11abiTRUE-cp312-cp312-linux_x86_64.whl", hash = "sha256:8e81f8c76435ad31aa41edc6c0c9d26de971a293b190183d2816da71547614d7" },
]
[package.metadata]
requires-dist = [
{ name = "ninja" },
{ name = "packaging" },
{ name = "torch" },
]
[[package]]
name = "certifi"
version = "2026.7.22"
@@ -106,6 +126,15 @@ wheels = [
{ url = "https://files.pythonhosted.org/packages/98/59/239c7259e669c46ddbcac0aa60e3a0ef00bfeaaa687f905b24dd6a7a10fe/cuda_pathfinder-1.8.2-py3-none-any.whl", hash = "sha256:4e65059febdb4d19d5cbc4798677e19db2b582f2f702f457b609e571690d357e", size = 62551 },
]
[[package]]
name = "einops"
version = "0.8.2"
source = { registry = "https://pypi.org/simple" }
sdist = { url = "https://files.pythonhosted.org/packages/2c/77/850bef8d72ffb9219f0b1aac23fbc1bf7d038ee6ea666f331fa273031aa2/einops-0.8.2.tar.gz", hash = "sha256:609da665570e5e265e27283aab09e7f279ade90c4f01bcfca111f3d3e13f2827", size = 56261 }
wheels = [
{ url = "https://files.pythonhosted.org/packages/2a/09/f8d8f8f31e4483c10a906437b4ce31bdf3d6d417b73fe33f1a8b59e34228/einops-0.8.2-py3-none-any.whl", hash = "sha256:54058201ac7087911181bfec4af6091bb59380360f069276601256a76af08193", size = 65638 },
]
[[package]]
name = "fastapi"
version = "0.118.0"
@@ -129,6 +158,31 @@ wheels = [
{ url = "https://files.pythonhosted.org/packages/bc/ac/8b9c6dc2aa7e9cf5582e20a337006c3428c3d752c2255dbaca2c78a8be45/filelock-4.0.4-py3-none-any.whl", hash = "sha256:0df72be195ca7892216d16f2edce8d9b93a571f02402972020a8cff84c594c7b", size = 108629 },
]
[[package]]
name = "fla-core"
version = "0.5.2"
source = { registry = "https://pypi.org/simple" }
dependencies = [
{ name = "einops" },
]
sdist = { url = "https://files.pythonhosted.org/packages/94/85/20bdc0fbbbaeec27b5b198e0dbbbc51161a267066a4aebd9e4b413637c6c/fla_core-0.5.2.tar.gz", hash = "sha256:9360fc412f784c1c8f05c320ec2902d5d968764178c9b8a92efc919e17a39680", size = 587835 }
wheels = [
{ url = "https://files.pythonhosted.org/packages/2d/ed/dfe19c4da779957eb6a42a26812f9b4e2280bf757a17a71933ff59ffcb98/fla_core-0.5.2-py3-none-any.whl", hash = "sha256:5e830c85bad3d0d34677f98ac7074d08687a3756f0f0499d95ceb96eb6920761", size = 819225 },
]
[[package]]
name = "flash-linear-attention"
version = "0.5.2"
source = { registry = "https://pypi.org/simple" }
dependencies = [
{ name = "fla-core" },
{ name = "transformers" },
]
sdist = { url = "https://files.pythonhosted.org/packages/d2/76/c180949eae5161b9fcf4c928cab52f0c257ce84c1a59b4462b1d2b6e5883/flash_linear_attention-0.5.2.tar.gz", hash = "sha256:c053d3a75c8f5b725063f719ae6bce0f4d9dac6da32735be8ae7d485a6c9820f", size = 208110 }
wheels = [
{ url = "https://files.pythonhosted.org/packages/90/d2/2070e3cf2148c5cce99ca4876633c4b7b89ec085323600bd3d47aeacd306/flash_linear_attention-0.5.2-py3-none-any.whl", hash = "sha256:dcf405d81f5426393b59037097aa700d0f4a841465d5028d5aa543f4502f2400", size = 399590 },
]
[[package]]
name = "fsspec"
version = "2026.9.0"
@@ -351,6 +405,24 @@ wheels = [
{ url = "https://files.pythonhosted.org/packages/7e/cd/fe58041e9011f307c490e3e17dd48cc516448f7c698a3f2d9d9d65d7e6a8/networkx-3.7-py3-none-any.whl", hash = "sha256:e3fd2c13a7814cee3746340d8d7f8598a67f16a58bf47fb7f8793fab6efca1b0", size = 2142205 },
]
[[package]]
name = "ninja"
version = "1.13.2"
source = { registry = "https://pypi.org/simple" }
sdist = { url = "https://files.pythonhosted.org/packages/ac/92/410b7917d16ab54c04b05cc32b9284803671d91cf79d33be6009c28d4ea8/ninja-1.13.2.tar.gz", hash = "sha256:525bfa3fc88aa30a4467df270fd5be6f9fcae8061d54d4df74ea1dc5abd5a975", size = 243739 }
wheels = [
{ url = "https://files.pythonhosted.org/packages/48/23/fcbe234a66966e35928c47b86336f92a7612db4781665f4e5f5fddef9630/ninja-1.13.2-py3-none-manylinux2014_ppc64le.manylinux_2_17_ppc64le.whl", hash = "sha256:227cbc3ae3e5e429692388103cae8c09451df086cd2d342dae0795af0d162547", size = 197676 },
{ url = "https://files.pythonhosted.org/packages/24/eb/a6ca97ef0ff7bb8bdcb395ec65a716e65d7c1f40896c3afe0090bb3e1535/ninja-1.13.2-py3-none-manylinux2014_s390x.manylinux_2_17_s390x.whl", hash = "sha256:1684c60d031c54c1d049541b64243c0c567dca5463dbd77682a8901780af293d", size = 187980 },
{ url = "https://files.pythonhosted.org/packages/6e/53/ebfed7b689c338dd8ebeec9c0730c8d56821292f14e2536e5f3ef1a05744/ninja-1.13.2-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:65a24341b5ac09fcadcc37082660be40a94174e51a937fabf6e2cae26225fa2c", size = 183365 },
{ url = "https://files.pythonhosted.org/packages/c7/d6/dcf06d7ab44ade992ae5aa1228feff317684b463a1bd47e8642b30ac922e/ninja-1.13.2-py3-none-manylinux_2_28_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:aa3d2ae5706a2c4d1e93edc951d1c6cbb45107413c404f8fde1741239efbc9a0", size = 155089 },
{ url = "https://files.pythonhosted.org/packages/e1/6b/6513c09c33382b17c05b4349b8e81437b18680d0d7ec6fb8f7edc29adda1/ninja-1.13.2-py3-none-manylinux_2_31_riscv64.whl", hash = "sha256:919572cbc3f233261ecd41fe1f3efc9d44aa02464a4588867e06a8b4f6f416ea", size = 152149 },
{ url = "https://files.pythonhosted.org/packages/37/04/c8c2dc5b2f5fee79a1691d490256b178b7e1af97d56769117422ae8a23cc/ninja-1.13.2-py3-none-musllinux_1_2_armv7l.whl", hash = "sha256:59d71c3e15b6b6f3d903eb0c27285544e0747ca59925ada7037bb1af781ad4b3", size = 466268 },
{ url = "https://files.pythonhosted.org/packages/10/a2/d8eedd25d0ae80b9e874aea362013e67416877c8f732eb4d5e7c971fdb9c/ninja-1.13.2-py3-none-musllinux_1_2_ppc64le.whl", hash = "sha256:b2f687437fac460b27b7eadc99039b1163016fb4ba7276e2782a192d9f24ee0e", size = 610806 },
{ url = "https://files.pythonhosted.org/packages/14/0f/696d96821fad1b5767fd311c1569dde8881a57412369bfe7b11bcbfde036/ninja-1.13.2-py3-none-musllinux_1_2_riscv64.whl", hash = "sha256:09de9ab04f7352f51570c73fd4913acb1e6c24be0a72cd8b80243d4d3ed04925", size = 533978 },
{ url = "https://files.pythonhosted.org/packages/5d/69/28844ca579156776a202217a7cd66f60d06a0710a935e879bb89ce396ecc/ninja-1.13.2-py3-none-musllinux_1_2_s390x.whl", hash = "sha256:6a87bf42b123abe2f37737300185f0a303a891899da85d73a3613ee80547e578", size = 653822 },
{ url = "https://files.pythonhosted.org/packages/f5/5f/c511f2952f94ab2966d60edd9c34e744ea32f2724b1184b62270bde55b3a/ninja-1.13.2-py3-none-musllinux_1_2_x86_64.whl", hash = "sha256:915bd482c4be41c75120fd67a22e0bb3f0fbb3bbc5f95b89787deadd59e27ef2", size = 544460 },
]
[[package]]
name = "numpy"
version = "2.2.6"
@@ -919,7 +991,7 @@ dependencies = [
[[package]]
name = "semif-serve"
version = "0.1.2"
version = "0.1.3"
source = { editable = "." }
dependencies = [
{ name = "fastapi" },
@@ -927,6 +999,10 @@ dependencies = [
]
[package.optional-dependencies]
fast = [
{ name = "causal-conv1d", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
{ name = "flash-linear-attention" },
]
model = [
{ name = "semif-phase1" },
{ name = "torch" },
@@ -940,12 +1016,14 @@ dev = [
[package.metadata]
requires-dist = [
{ name = "causal-conv1d", marker = "platform_machine == 'x86_64' and sys_platform == 'linux' and extra == 'fast'", url = "https://github.com/Dao-AILab/causal-conv1d/releases/download/v1.7.0/causal_conv1d-1.7.0+cu12torch2.10cxx11abiTRUE-cp312-cp312-linux_x86_64.whl" },
{ name = "fastapi", specifier = "==0.118.0" },
{ name = "flash-linear-attention", marker = "extra == 'fast'", specifier = "==0.5.2" },
{ name = "semif-phase1", marker = "extra == 'model'", git = "https://github.com/TheoLeeCJ/SemIf-OpenJev?rev=23cf1f39fc9534fe81437200959b6dfc7106e45a" },
{ name = "torch", marker = "extra == 'model'", specifier = "==2.10.0", index = "https://download.pytorch.org/whl/cu128" },
{ name = "uvicorn", specifier = "==0.37.0" },
]
provides-extras = ["model"]
provides-extras = ["model", "fast"]
[package.metadata.requires-dev]
dev = [