mirror of
https://github.com/Niko1221/Strata.git
synced 2026-11-07 18:35:04 -10:00
Version 0.1.40.3: MIN_ENGINE 0.1.40.3; Intel Arc setup text says supported (B70, A750) instead of 'no Intel card here'
This commit is contained in:
Niko1221 committed
1 parent
34ff00bdd0
commit
d5ea713374
5 files changed
+10
-10
No files matched your search
+1
-1
@@ -8,7 +8,7 @@
|
||||
# below refuses anything older: silently building for a pre-Ampere arch would produce a binary that runs
|
||||
# and is wrong about its own performance, which is the worst of both outcomes.
|
||||
cmake_minimum_required(VERSION 3.24)
|
||||
project(strata VERSION 0.1.40.2 LANGUAGES CXX)
|
||||
project(strata VERSION 0.1.40.3 LANGUAGES CXX)
|
||||
# EXPERIMENTAL: the Intel Arc engine, the SYCL port in sycl/ (PR #423; icpx + oneMKL, docs/INTEL_ARC.md). It is its
|
||||
# own project that takes the unported sources from src/; with this on, nothing below is configured.
|
||||
option(STRATA_ENABLE_SYCL "EXPERIMENTAL: build the Intel Arc (SYCL) engine from sycl/ - docs/INTEL_ARC.md" OFF)
|
||||
|
||||
+1
-2
@@ -21,8 +21,7 @@ There is **no ready-made Intel engine** in the release zips. You build it from s
|
||||
| Community | Arc B580 12 GB, WSL2 | IQ2_XS ~21 tok/s, Coder ~15 tok/s; **device loss also seen** | #423 |
|
||||
| Community | 2x Arc Pro B60 24 GB, `--layer-split` | Coder IQ1_M: 56.6 tok/s decode, 428 tok/s prompt (2,129 tokens); Flash-Next IQ2_XS: 58.6-61.3 tok/s decode, 436 tok/s prompt; all experts in VRAM; 8K context | [bench/results/2026-10-04-community-2x-arc-pro-b60](../bench/results/2026-10-04-community-2x-arc-pro-b60/README.md) |
|
||||
| Community | one Arc Pro B60 24 GB | Coder IQ1_M with 4,042 of 12,288 experts mirrored in RAM: 11.9 tok/s decode (needs the ring-wait fix from the same report) | same |
|
||||
| Strata maintainers | no Arc | compile check and kernel tests on a CPU device only (below) | this release |
|
||||
| Strata maintainers (0.1.41 work) | Arc Pro B70 32 GB (xe) and Arc A750 8 GB (i915), test machines | Flash-Next IQ3_S on the B70 with 11.5k of 24.6k experts resident: 10-prompt gate passes, prompt 980-1,000 tok/s on 4K tokens, decode 30 (prose) to 41 (code) tok/s; the xe aliased-pages fix and the `NO_HOST` requirement are in [INTEL.md](INTEL.md) | INTEL.md, "Arc Pro B70 with a model that does not fit" |
|
||||
| Strata maintainers (0.1.40.2 and later) | Arc Pro B70 32 GB (xe) and Arc A750 8 GB (i915), test machines | Flash-Next IQ3_S on the B70 with 11.5k of 24.6k experts resident: 10-prompt gate passes, prompt 980-1,000 tok/s on 4K tokens, decode 30 (prose) to 41 (code) tok/s; the xe aliased-pages fix and the `NO_HOST` requirement are in [INTEL.md](INTEL.md) | INTEL.md, "Arc Pro B70 with a model that does not fit" |
|
||||
|
||||
The 0.1.39 port re-migrates 0.1.38's port onto the 0.1.39 engine sources (the #606 NaN fix, the #649 verify
|
||||
trace, the new prompt paths). 0.1.39's `sycl/` did not compile against 0.1.39's own engine sources (#784: the
|
||||
|
||||
@@ -192,7 +192,7 @@ CUDA12_WHEELS = ["nvidia-cublas-cu12==12.9.1.4", "nvidia-cuda-runtime-cu12==12.9
|
||||
# toolkit the CUDA 12 zip is built with (cuBLAS 12.9.1.4, runtime 12.9.79). Not tested on such an old driver here.
|
||||
CUDA12_MIN_DRIVER = 528 if WIN else 525
|
||||
ENGINE12_DIR = "engine-cuda12"
|
||||
MIN_ENGINE = (0, 1, 40, 2) # versions compare all four numbers; v0.1.40.2: F4 verify windows, the Linux file tier (#1194), the stager wait (#1057), #1264/#1201/#1139 fixes, opt-in CPU share (#1282), Intel Arc; v0.1.40: --resident-experts on a layer split with the split+resident variant (#848), --kv k8v4 with KV streaming (#711); v0.1.39: the #577 file-tier regression fixed, the OpenAI Responses API (#451, Codex), a reply stuck on one token ended (#606), the head before the arena (#620), effort_position (#458), --vram-reserve hot resize opt-in (#533), PR batch; v0.1.38: prompts faster (one gather per expert group #372, the first chunk's PLE rows beside layer 0 #374, DeltaNet three heads per thread #413), --kv q4_0 prompts on tensor cores (#452), Q5_0 experts on the GPU (#473), IQ4_XS on AVX-2 (#415), unbuffered expert loading on Windows (#357 #362), --peer-device (#531), a 6 GB card starts (#496), PR batch; v0.1.37: a silent engine is restarted (#481), Windows AMD counts the desktop's VRAM (#380 #377 #497), a steadier PCIe probe (#485), fixes #496 #495 #498 #505 #493; v0.1.36: a cancelled prompt logged as read so far (#471), the draft-head hint (#474), UPDATE.bat (#475), --expert-profile-save (#477); v0.1.35: Windows AMD uses its bundled HIP runtime (#468 #461), the low-RAM resident mode on Windows 32 GB (#467), fixes #460 #459 #446 #447 #457 #448 #444; v0.1.34: AMD on Windows (a ready-made HIP engine), an MCP server for AI assistants (tools/strata_mcp.py), a shorter README; v0.1.33: a portable image encoder again (#411 #412), setup recommends instead of forcing (#406 #403 #364 #384), fixes #352 #365 #369 #371 #375 #393 #408 #414; v0.1.32: split prompts faster (#340), AMD router +12%, Unsloth Q4 in setup, faster Q4 prompts, #326/#327/#342/#344 fixes, PR batch; v0.1.31: Unsloth UD-Q4_K_XL (experimental), GGUF-in-place low-RAM mode, Windows GGUF load 2x, server race + tokenizer fixes, AMD intrinsics; v0.1.30: short prompts faster (streaming from 1024 tokens), resident low-RAM variant, multi-GPU session carve, RDNA4; v0.1.29: sampled answers faster (split top-k), #154 correctness fixes; v0.1.28: the expert cache reserves the draft head, a cancelled request no longer fails the next; v0.1.27: RTX 20 (sm_75) in the ready-made engine, the HIP build without CUDA headers; v0.1.26: the draft layer's prompt pass in batches; v0.1.25: faster prompts (grouping off the copy engine, fused hyper-connection kernels), AMD HIP backend, --kv k8v4; v0.1.24: long prompts faster (QSA select on tensor cores); v0.1.23: image requests honor sampling, 8 GB cards start, batched verify window; v0.1.22: faster prompts (tensor-core attention), multi-GPU across images/steering/KV streaming; v0.1.21: multi-GPU layer split (--gpus); v0.1.20: system-prompt checkpoint, PCIe probe, hit rate; v0.1.19: penalties
|
||||
MIN_ENGINE = (0, 1, 40, 3) # versions compare all four numbers; v0.1.40.3: the #1357 MTP router guard, the #1376 Windows HIP cache floor, #461 runtime DLLs, Intel A750 first-request fix; v0.1.40.2: F4 verify windows, the Linux file tier (#1194), the stager wait (#1057), #1264/#1201/#1139 fixes, opt-in CPU share (#1282), Intel Arc; v0.1.40: --resident-experts on a layer split with the split+resident variant (#848), --kv k8v4 with KV streaming (#711); v0.1.39: the #577 file-tier regression fixed, the OpenAI Responses API (#451, Codex), a reply stuck on one token ended (#606), the head before the arena (#620), effort_position (#458), --vram-reserve hot resize opt-in (#533), PR batch; v0.1.38: prompts faster (one gather per expert group #372, the first chunk's PLE rows beside layer 0 #374, DeltaNet three heads per thread #413), --kv q4_0 prompts on tensor cores (#452), Q5_0 experts on the GPU (#473), IQ4_XS on AVX-2 (#415), unbuffered expert loading on Windows (#357 #362), --peer-device (#531), a 6 GB card starts (#496), PR batch; v0.1.37: a silent engine is restarted (#481), Windows AMD counts the desktop's VRAM (#380 #377 #497), a steadier PCIe probe (#485), fixes #496 #495 #498 #505 #493; v0.1.36: a cancelled prompt logged as read so far (#471), the draft-head hint (#474), UPDATE.bat (#475), --expert-profile-save (#477); v0.1.35: Windows AMD uses its bundled HIP runtime (#468 #461), the low-RAM resident mode on Windows 32 GB (#467), fixes #460 #459 #446 #447 #457 #448 #444; v0.1.34: AMD on Windows (a ready-made HIP engine), an MCP server for AI assistants (tools/strata_mcp.py), a shorter README; v0.1.33: a portable image encoder again (#411 #412), setup recommends instead of forcing (#406 #403 #364 #384), fixes #352 #365 #369 #371 #375 #393 #408 #414; v0.1.32: split prompts faster (#340), AMD router +12%, Unsloth Q4 in setup, faster Q4 prompts, #326/#327/#342/#344 fixes, PR batch; v0.1.31: Unsloth UD-Q4_K_XL (experimental), GGUF-in-place low-RAM mode, Windows GGUF load 2x, server race + tokenizer fixes, AMD intrinsics; v0.1.30: short prompts faster (streaming from 1024 tokens), resident low-RAM variant, multi-GPU session carve, RDNA4; v0.1.29: sampled answers faster (split top-k), #154 correctness fixes; v0.1.28: the expert cache reserves the draft head, a cancelled request no longer fails the next; v0.1.27: RTX 20 (sm_75) in the ready-made engine, the HIP build without CUDA headers; v0.1.26: the draft layer's prompt pass in batches; v0.1.25: faster prompts (grouping off the copy engine, fused hyper-connection kernels), AMD HIP backend, --kv k8v4; v0.1.24: long prompts faster (QSA select on tensor cores); v0.1.23: image requests honor sampling, 8 GB cards start, batched verify window; v0.1.22: faster prompts (tensor-core attention), multi-GPU across images/steering/KV streaming; v0.1.21: multi-GPU layer split (--gpus); v0.1.20: system-prompt checkpoint, PCIe probe, hit rate; v0.1.19: penalties
|
||||
# KV bytes per context token and attention layer: 8-bit 1056, rotated 4-bit 576, hybrid K8V4 (8-bit K, 4-bit V) 816
|
||||
KV_CELL_BYTES = {"q4_0": 576, "k8v4": 816}
|
||||
PY_PACKAGES = ["numpy", "jinja2", "regex", "pyyaml", "tqdm", "requests", "cmake", "ninja", "pillow", "psutil"]
|
||||
@@ -4489,8 +4489,8 @@ def sycl_setup(argv) -> int:
|
||||
Intel engine: it is compiled from source on the PC (docs/INTEL_ARC.md), then sycl/setup_intel.py runs this setup
|
||||
with the Intel steps swapped in. Nothing of the CUDA / HIP paths is used or changed."""
|
||||
say()
|
||||
warn("Intel Arc (--backend sycl) is EXPERIMENTAL: a community port of the engine, not tested by the Strata "
|
||||
"maintainers (no Intel card here). Expect rough edges; issues with your card and driver versions help.")
|
||||
say(" Intel Arc (--backend sycl): supported since 0.1.40.2 on Linux, tested on an Arc Pro B70 (xe) and an Arc A750 "
|
||||
"(i915); other Arc cards and driver versions are untested, and reports help (docs/INTEL.md).")
|
||||
if WIN:
|
||||
fail("the Intel Arc engine has no Windows setup yet (no ready-made Intel engine either)",
|
||||
"run it on Linux (Ubuntu 24.04 with Intel's GPU driver and oneAPI): docs/INTEL_ARC.md")
|
||||
|
||||
@@ -20,8 +20,9 @@ import setup # noqa: E402
|
||||
|
||||
class HotfixTag(unittest.TestCase):
|
||||
def test_a_four_part_version_compares_all_four_numbers(self):
|
||||
ver = tuple(int(x) for x in "0.1.40.2".split(".")[:4] if x.isdigit())
|
||||
self.assertEqual(ver, (0, 1, 40, 2))
|
||||
ver = tuple(int(x) for x in ".".join(map(str, setup.MIN_ENGINE)).split(".")[:4] if x.isdigit())
|
||||
self.assertEqual(ver, setup.MIN_ENGINE)
|
||||
self.assertEqual(len(ver), 4)
|
||||
self.assertGreaterEqual(ver, setup.MIN_ENGINE)
|
||||
self.assertLess((0, 1, 40), setup.MIN_ENGINE) # an installed 0.1.40 engine is replaced
|
||||
|
||||
|
||||
@@ -34,14 +34,14 @@ class SyclBackend(unittest.TestCase):
|
||||
def test_windows_stops_with_the_docs_pointer(self):
|
||||
rc, out, call = run(["--backend", "sycl"], win=True)
|
||||
self.assertEqual(rc, ("exit", 1))
|
||||
self.assertIn("EXPERIMENTAL", out)
|
||||
self.assertIn("supported since 0.1.40.2", out)
|
||||
self.assertIn("docs/INTEL_ARC.md", out)
|
||||
call.assert_not_called()
|
||||
|
||||
def test_linux_hands_over_to_setup_intel_without_the_backend_flag(self):
|
||||
rc, out, call = run(["--model", "IQ2_XS", "--backend", "sycl", "--port", "8085"], win=False)
|
||||
self.assertEqual(rc, 0)
|
||||
self.assertIn("EXPERIMENTAL", out)
|
||||
self.assertIn("supported since 0.1.40.2", out)
|
||||
self.assertIn("built from source", out)
|
||||
cmd = call.call_args[0][0]
|
||||
self.assertTrue(cmd[1].endswith(str(Path("sycl") / "setup_intel.py")))
|
||||
|
||||
Reference in new issue
Block a user