Skip to content

Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.

C++

Gyro on Intel Arc: README and MODELS say so, ctest cases for the trellis kernels, setup's Gyro memory warning reads the real RAM on the Intel path

Niko122179946b9c4fc4
diff --git a/README.de.md b/README.de.md
index 1bfd321d5..72eb048e4 100644
--- a/README.de.md
+++ b/README.de.md
@@ -164,7 +164,7 @@ Größen, mehr oder weniger stark komprimiert. Kleinere Größen sind schneller.
| **Gyro-S** | 27,6 GB | 58,5 GB | Für PCs, auf denen nichts anderes läuft. Die Qualität ist geringer als bei IQ2_XS. |
| **Gyro-M** | 35 GB | 92,0 GB | Ungefähr die Qualität von IQ3_XXS bei etwa 20 % weniger GPU-Speicher. Es schreibt Antworten auf einer 12-GB-Karte und auf einer R9700 17-21 % langsamer als Q2_0. |
-Gyro läuft auf NVIDIA- und AMD-Karten (bei AMD experimentell), nicht auf Intel Arc und nicht auf dem Mac. Details und Messwerte:
+Gyro läuft auf NVIDIA- und AMD-Karten (bei AMD experimentell) und auf Intel-Arc-Karten unter Linux (getestet auf einer Arc A750), nicht auf dem Mac. Details und Messwerte:
[docs/GYRO.md](docs/GYRO.md). Starte es mit `strata launch --model gyro-s`.
### Die anderen Größen
diff --git a/README.es.md b/README.es.md
index 679fafd5a..0be4b744f 100644
--- a/README.es.md
+++ b/README.es.md
@@ -164,7 +164,7 @@ comprimido. Los tamaños pequeños son más rápidos. Los grandes son algo más
| **Gyro-S** | 27,6 GB | 58,5 GB | Para PC donde no cabe nada más. La calidad es menor que la de IQ2_XS. |
| **Gyro-M** | 35 GB | 92,0 GB | Calidad parecida a IQ3_XXS con un 20 % menos de memoria de GPU. Escribe las respuestas un 17-21 % más despacio que Q2_0 en una tarjeta de 12 GB y en una R9700. |
-Gyro funciona en tarjetas NVIDIA y AMD (en AMD es experimental), no en Intel Arc ni en el Mac. Detalles y mediciones:
+Gyro funciona en tarjetas NVIDIA y AMD (en AMD es experimental) y en tarjetas Intel Arc con Linux (probado en una Arc A750), no en el Mac. Detalles y mediciones:
[docs/GYRO.md](docs/GYRO.md). Arráncalo con `strata launch --model gyro-s`.
### Los demás tamaños
diff --git a/README.fr.md b/README.fr.md
index 9ce7d44e1..085648539 100644
--- a/README.fr.md
+++ b/README.fr.md
@@ -165,7 +165,7 @@ plus ou moins compressé. Les petites tailles sont plus rapides. Les grandes son
| **Gyro-S** | 27,6 Go | 58,5 Go | Pour les PC où rien d'autre ne tient. La qualité est inférieure à celle d'IQ2_XS. |
| **Gyro-M** | 35 Go | 92,0 Go | À peu près la qualité d'IQ3_XXS avec environ 20 % de mémoire GPU en moins. Il écrit les réponses 17 à 21 % plus lentement que Q2_0 sur une carte de 12 Go et sur une R9700. |
-Gyro fonctionne sur les cartes NVIDIA et AMD (expérimental sur AMD), pas sur Intel Arc ni sur Mac. Détails et mesures :
+Gyro fonctionne sur les cartes NVIDIA et AMD (expérimental sur AMD) et sur les cartes Intel Arc sous Linux (testé sur une Arc A750), pas sur Mac. Détails et mesures :
[docs/GYRO.md](docs/GYRO.md). Lancez-le avec `strata launch --model gyro-s`.
### Les autres tailles
diff --git a/README.ja.md b/README.ja.md
index e3d7e8353..8402120f0 100644
--- a/README.ja.md
+++ b/README.ja.md
@@ -158,7 +158,7 @@ AI ツールは [MCP サーバー](docs/MCP_SERVER.md)を通じて、Strata の
| **Gyro-S** | 27.6 GB | 58.5 GB | ほかのサイズが入らないPC向け。品質は IQ2_XS より下がります。 |
| **Gyro-M** | 35 GB | 92.0 GB | 品質は IQ3_XXS 程度で、GPUメモリの使用量は約20%少なくなります。12 GBのカードと R9700 では、回答の書き出しが Q2_0 より17〜21%遅くなります。 |
-Gyro は NVIDIA と AMD のカードで動きます(AMD は実験的)。Intel Arc と Mac では動きません。詳細と測定結果は
+Gyro は NVIDIA と AMD のカード(AMD は実験的)、および Linux 上の Intel Arc カード(Arc A750 で確認)で動きます。Mac では動きません。詳細と測定結果は
[docs/GYRO.md](docs/GYRO.md) にあります。`strata launch --model gyro-s` で起動します。
### ほかのサイズ
diff --git a/README.md b/README.md
index 4b7b92e40..394b1a3c6 100644
--- a/README.md
+++ b/README.md
@@ -159,7 +159,7 @@ Smaller sizes are faster. Larger sizes are a bit smarter.
| **Gyro-S** | 27.6 GB | 58.5 GB | For PCs that fit nothing else. Quality is lower than IQ2_XS. |
| **Gyro-M** | 35 GB | 92.0 GB | About IQ3_XXS quality with about 20% less GPU memory. It writes answers 17-21% slower than Q2_0 on a 12 GB card and on an R9700. |
-Gyro runs on NVIDIA and AMD cards (AMD is experimental), not on Intel Arc or the Mac. Details and measurements:
+Gyro runs on NVIDIA and AMD cards (AMD is experimental) and on Intel Arc cards on Linux (tested on an Arc A750), not on the Mac. Details and measurements:
[docs/GYRO.md](docs/GYRO.md). Start it with `strata launch --model gyro-s`.
### The other sizes
diff --git a/README.pt-BR.md b/README.pt-BR.md
index d97e534fb..9e34e5aa9 100644
--- a/README.pt-BR.md
+++ b/README.pt-BR.md
@@ -158,7 +158,7 @@ O [Gyro](https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF) (da age
| **Gyro-S** | 27,6 GB | 58,5 GB | Para PCs onde nenhum outro tamanho cabe. A qualidade é menor que a do IQ2_XS. |
| **Gyro-M** | 35 GB | 92,0 GB | Qualidade próxima à do IQ3_XXS, com cerca de 20% menos memória da GPU. Ele escreve as respostas 17 a 21% mais devagar que o Q2_0 numa placa de 12 GB e numa R9700. |
-O Gyro roda em placas NVIDIA e AMD (AMD é experimental), e não roda em Intel Arc nem no Mac. Detalhes e medições:
+O Gyro roda em placas NVIDIA e AMD (AMD é experimental) e em placas Intel Arc no Linux (testado numa Arc A750), e não roda no Mac. Detalhes e medições:
[docs/GYRO.md](docs/GYRO.md). Inicie com `strata launch --model gyro-s`.
### Os outros tamanhos
diff --git a/README.zh-CN.md b/README.zh-CN.md
index d16844858..72ec629d7 100644
--- a/README.zh-CN.md
+++ b/README.zh-CN.md
@@ -157,7 +157,7 @@ Set up Strata on this PC for me: https://github.com/Niko1221/Strata - follow doc
| **Gyro-S** | 27.6 GB | 58.5 GB | 给其他尺寸都放不下的电脑。质量低于 IQ2_XS。 |
| **Gyro-M** | 35 GB | 92.0 GB | 质量约相当于 IQ3_XXS,显存占用少约 20%。在 12 GB 显卡和 R9700 上,输出回答比 Q2_0 慢 17-21%。 |
-Gyro 可在 NVIDIA 和 AMD 显卡上运行(AMD 为实验性支持),不支持 Intel Arc 和 Mac。详情和测试数据:
+Gyro 可在 NVIDIA 和 AMD 显卡(AMD 为实验性支持)以及 Linux 上的 Intel Arc 显卡(已在 Arc A750 上测试)上运行,不支持 Mac。详情和测试数据:
[docs/GYRO.md](docs/GYRO.md)。用 `strata launch --model gyro-s` 启动。
### 其他尺寸
diff --git a/docs/MODELS.md b/docs/MODELS.md
index c51fae258..a828be85d 100644
--- a/docs/MODELS.md
+++ b/docs/MODELS.md
@@ -19,7 +19,7 @@ other size stays a choice. A PC with a card of 12 GB and 32 GB of RAM is in the
| Your memory (RAM + VRAM) | Pre-selected | Notes |
| --- | --- | --- |
| **32 GB** | **Gyro-S**, or the **Coder** for code | The Coder needs 32 GB of RAM and keeps half of the experts, so it is weaker outside code ([why](#coder)). Gyro-S has lower quality than the 2-3-bit sizes, and the 2-bit Q2_0 decodes faster |
-| **40 GB** | **Swift 1.5** IQ2_XS (39.2 GB) | Gyro-M (35 GB) is the other choice: close to IQ3_XXS in quality, but it writes answers 17-21% slower than Q2_0. Gyro runs on NVIDIA and AMD cards, not on Intel Arc or the Mac |
+| **40 GB** | **Swift 1.5** IQ2_XS (39.2 GB) | Gyro-M (35 GB) is the other choice: close to IQ3_XXS in quality, but it writes answers 17-21% slower than Q2_0. Gyro runs on NVIDIA, AMD and Intel Arc cards, not on the Mac |
| **48 GB** | **IQ3_XXS** (47.0 GB) | Slower, better |
| **64 GB** | **IQ3_S** (54.8 GB), the original. Swift 1.5's IQ3_S is an equal choice | The slowest, the best: it matches the full model on the published tests |
| **96 GB or more** | [Unsloth's UD-IQ4_XS to Q8_0](#unsloths-files-ud-iq4_xs-to-q8_0), 4 to 8 bit | UD-IQ4_XS keeps every expert in RAM from about 80 GB of RAM, Q8_0 from about 136 GB. Up to 192 GB, the 5-8-bit files fit in RAM. Setup still pre-selects IQ3_S here, except on a Strix Halo |
@@ -183,7 +183,7 @@ experts in a rotated basis. They are the smallest options and never setup's defa
| **Gyro-S** | 58.5 GB | 27.6 GB | the smallest option, for PCs with little memory; lower quality |
| **Gyro-M** | 92.0 GB | 35 GB | about IQ3_XXS quality with about 20% less GPU memory; decodes about 20% slower than Q2_0 on 12 GB cards |
-NVIDIA, and AMD (experimental); not Intel Arc or the Mac. Gyro reads pictures (the same rule as above). It needs
+Gyro runs on NVIDIA and AMD (experimental) cards and on Intel Arc cards on Linux (tested on an Arc A750; the Arc engine is built from this release's source, see [INTEL.md](INTEL.md)); not on the Mac. Gyro reads pictures (the same rule as above). On NVIDIA and AMD it needs
engine 0.1.45.1 or newer; if the engine is older, the app says "This model needs a newer engine" and offers Update
engine. Gyro gives correct output on RTX 50 cards with engine 0.1.45.3. Details, measurements and credit:
[GYRO.md](GYRO.md).
diff --git a/sycl/CMakeLists.txt b/sycl/CMakeLists.txt
index 4dc7ae000..6f963d098 100644
--- a/sycl/CMakeLists.txt
+++ b/sycl/CMakeLists.txt
@@ -429,6 +429,10 @@ if(STRATA_SYCL_PARITY)
target_include_directories(tq_sycl_parity PRIVATE ${STRATA_ROOT}/third_party/ggml)
target_compile_options(tq_sycl_parity PRIVATE -ffp-contract=off)
target_link_libraries(tq_sycl_parity PRIVATE strata_engine strata_kernels strata_kernels_cpu)
+ # random trellis codes (every bit pattern is valid): the GPU kernels against a host replay of their own arithmetic, bit for bit,
+ # the prompt-path dequantizers against the fp16 codebook, and the Hadamard rotations against the CPU's
+ add_test(NAME tq_sycl_parity_synthetic COMMAND tq_sycl_parity --synthetic 145/146 144/144 144/146 145/144)
+ add_test(NAME tq_sycl_parity_synthetic_hadamard COMMAND tq_sycl_parity --synthetic --hadamard 145/146 144/144)
# MODEL-UPDATE: the CPU pool's expert rows per format against the RAM bandwidth (as the top-level build's)
add_executable(kq_cpu_bench ${STRATA_ROOT}/src/kernels/cpu/kq_cpu_bench.cpp)
target_link_libraries(kq_cpu_bench PRIVATE strata_kernels_cpu)
diff --git a/sycl/setup_intel.py b/sycl/setup_intel.py
index c4395e424..d7b66829c 100644
--- a/sycl/setup_intel.py
+++ b/sycl/setup_intel.py
@@ -88,6 +88,12 @@ def intel_gpus():
return found
+def real_mem_have(mem_have, real_ram: float, fake_ram: float):
+ """setup.mem_have for the Intel path: the RAM it is given is the 1024 GB stand-in of setup's RAM rule, so a model that counts VRAM + RAM
+ (Gyro) would never warn; it gets the real RAM."""
+ return lambda model, ram, vram=0.0: mem_have(model, real_ram if ram >= fake_ram else ram, vram)
+
+
def sycl_engine():
"""The SYCL build: (binary, why-not)."""
if S.WIN:
@@ -406,8 +412,7 @@ def install(argv) -> None:
# Gyro (mem_total) counts VRAM + RAM together and is the one family whose experts are NOT streamed from the file into VRAM: on
# an Arc they sit in the RAM arena and the CPU computes what the card does not hold. Its memory warning ("needs 27.6 GB of VRAM +
# RAM together, you have ...", never a refusal) therefore reads the real RAM, not the 1024 GB above.
- mem_have = S.mem_have
- S.mem_have = lambda model, ram, vram=0.0: mem_have(model, real_ram if ram >= fake_ram else ram, vram)
+ S.mem_have = real_mem_have(S.mem_have, real_ram, fake_ram)
bench_tips = S.bench_tips
diff --git a/tools/test_setup_sycl.py b/tools/test_setup_sycl.py
index 10ba23aa9..339243c85 100644
--- a/tools/test_setup_sycl.py
+++ b/tools/test_setup_sycl.py
@@ -206,6 +206,19 @@ class ToSycl(unittest.TestCase):
self.assertEqual(c["env"]["STRATA_VERIFY_NO_HOST"], "0")
self.assertEqual(c["env"]["STRATA_SYCL_BIN"], "build-sycl/strata")
+ def test_gyro_memory_warning_reads_the_real_ram(self):
+ # setup's RAM rule is fed 1024 GB on the Intel path; Gyro counts VRAM + RAM, so it must see the real RAM
+ m = self.mod()
+ import importlib.util
+ spec = importlib.util.spec_from_file_location("setup_for_test", ROOT / "setup.py")
+ s = importlib.util.module_from_spec(spec)
+ spec.loader.exec_module(s)
+ have = m.real_mem_have(s.mem_have, 16.0, 1024.0)
+ self.assertEqual(have("Gyro-S", 1024.0, 8.0), 24.0) # 16 GB RAM + 8 GB VRAM: below Gyro-S's 27.6, setup warns
+ self.assertLess(have("Gyro-S", 1024.0, 8.0), s.MODELS["Gyro-S"]["ram_gb"])
+ self.assertEqual(have("Gyro-S", 40.0, 8.0), 48.0) # a real RAM figure passes through
+ self.assertEqual(have("Q2_0", 1024.0, 8.0), 16.0) # the other sizes count RAM only, as before
+
def test_a_series_keeps_gyros_ngram_table_on_the_ssd(self):
# Gyro's table is half of the 58-92 GB file: --ple-io ram would not fit beside the experts' arena on a 64 GB PC
m = self.mod()