{"id":4237,"date":"2026-04-15T10:45:05","date_gmt":"2026-04-15T07:45:05","guid":{"rendered":"https:\/\/mojaz.ai\/llama-cpp-b8797-hexagon-hmx-optimizations-prebuilt-releases\/"},"modified":"2026-04-15T10:45:05","modified_gmt":"2026-04-15T07:45:05","slug":"llama-cpp-b8797-hexagon-hmx-optimizations-prebuilt-releases","status":"publish","type":"post","link":"https:\/\/mojaz.ai\/staging\/llama-cpp-b8797-hexagon-hmx-optimizations-prebuilt-releases\/","title":{"rendered":"llama.cpp b8797 \u064a\u062b\u0648\u0651\u0631 \u0623\u062f\u0627\u0621 Hexagon \u0628\u06406 \u062a\u062d\u0633\u064a\u0646\u0627\u062a HMX \u0648\u0627\u0633\u0639\u0629"},"content":{"rendered":"

\u0628\u0642\u0644\u0645: \u064a\u0648\u0633\u0641<\/a> | \u0645\u062d\u0631\u0631 \u0623\u062f\u0648\u0627\u062a \u0627\u0644\u0630\u0643\u0627\u0621 \u0627\u0644\u0627\u0635\u0637\u0646\u0627\u0639\u064a \u00b7 \u0635\u0648\u062a \u062a\u062d\u0631\u064a\u0631\u064a \u0628\u0625\u0634\u0631\u0627\u0641 \u0628\u0634\u0631\u064a<\/a><\/p>\n

\n

\u064a\u062f\u0634\u0646 \u0645\u0637\u0648\u0631\u0648 llama.cpp<\/strong> \u0627\u0644\u0646\u0633\u062e\u0629 b8797<\/strong> \u0628\u062a\u062d\u0648\u0644\u0627\u062a \u0647\u0646\u062f\u0633\u064a\u0629 \u0639\u0645\u064a\u0642\u0629 \u0641\u064a \u0643\u064a\u0641\u064a\u0629 \u062a\u0639\u0627\u0645\u0644 \u0627\u0644\u0645\u0643\u062a\u0628\u0629 \u0645\u0639 \u0645\u0639\u0627\u0644\u062c\u0627\u062a Hexagon<\/strong> \u0645\u0646 \u0643\u0648\u0627\u0644\u0643\u0648\u0645\u060c \u0645\u0633\u062a\u0628\u062f\u0644\u064a\u0646 \u0627\u0644\u0628\u0646\u064a\u0629 \u0627\u0644\u0645\u062a\u0632\u0627\u0645\u0646\u0629 \u0627\u0644\u0633\u0627\u0628\u0642\u0629 \u0628\u0646\u0638\u0627\u0645 hmx-queue<\/strong> \u064a\u062d\u0627\u0643\u064a \u0648\u0627\u062c\u0647\u0629 dma-queue \u0644\u062a\u062d\u0633\u064a\u0646 \u0627\u0644\u062a\u062f\u0627\u062e\u0644 \u0648\u0627\u0644\u0623\u062f\u0627\u0621.<\/p>\n

    \n
  1. hmx-worker \u0627\u0644\u062c\u062f\u064a\u062f:<\/strong> \u0646\u0638\u0627\u0645 \u062e\u064a\u0637 \u0645\u062e\u0635\u0635 \u0644\u0644\u062d\u0648\u0633\u0628\u0629 HMX \u064a\u062a\u062f\u0627\u062e\u0644 \u0645\u0639 \u0639\u0645\u0644\u064a\u0627\u062a HVX dequant\/DMA \u0641\u064a \u062e\u0637 \u0627\u0644\u0625\u0646\u062a\u0627\u062c\u060c \u0645\u064f\u0633\u062a\u0628\u062f\u0644\u0627\u064b \u0627\u0644\u0645\u0643\u0627\u0644\u0645\u0627\u062a \u0627\u0644\u0645\u062a\u0632\u0627\u0645\u0646\u0629 \u0627\u0644\u062a\u064a \u0643\u0627\u0646\u062a \u062a\u062d\u062c\u0628 \u0627\u0644\u062e\u064a\u0637 \u0627\u0644\u0631\u0626\u064a\u0633\u064a<\/li>\n
  2. \u0627\u0644\u0628\u062d\u062b \u0627\u0644\u0645\u0639\u062a\u0645\u062f \u0639\u0644\u0649 \u0627\u0644\u062a\u0643\u0644\u0641\u0629:<\/strong> \u062e\u0648\u0627\u0631\u0632\u0645\u064a\u0629 VTCM chunk search \u0645\u062d\u0633\u0651\u0646\u0629 \u0644\u0639\u0645\u0644\u064a\u0627\u062a out-stationary matmul \u062a\u062d\u062f\u062f \u0627\u0644\u062d\u062c\u0645 \u0627\u0644\u0623\u0645\u062b\u0644 \u0644\u0644\u0630\u0627\u0643\u0631\u0629 \u0627\u0644\u0645\u0624\u0642\u062a\u0629<\/li>\n
  3. \u0625\u0635\u0644\u0627\u062d futex race:<\/strong> \u062d\u0644 \u0645\u0634\u0643\u0644\u0629 \u0627\u0644\u0633\u0628\u0627\u0642 \u0641\u064a hmx_worker_drain \u0639\u0628\u0631 \u062d\u0641\u0638 \u0627\u0644\u0645\u062a\u063a\u064a\u0631 \u0627\u0644\u0628\u0648\u0644\u064a\u0646\u064a \u0645\u062d\u0644\u064a\u0627\u064b \u0644\u062a\u062c\u0646\u0628 \u0627\u0644\u062a\u062d\u0645\u064a\u0644 \u0627\u0644\u0630\u0631\u064a \u0627\u0644\u0645\u0636\u0627\u0639\u0641<\/li>\n
  4. \u062a\u062d\u0633\u064a\u0646 scatter\/transpose:<\/strong> \u0627\u0633\u062a\u062e\u062f\u0627\u0645 HMX intrinsics \u0627\u0644\u0623\u0635\u0644\u064a\u0629 \u0644\u0639\u0645\u0644\u064a\u0627\u062a \u0625\u0639\u0627\u062f\u0629 \u062a\u0631\u062a\u064a\u0628 \u0627\u0644\u0628\u064a\u0627\u0646\u0627\u062a \u0645\u0639 \u062a\u062d\u0633\u064a\u0646 \u0627\u0644\u0623\u062f\u0627\u0621<\/li>\n
  5. \u0636\u0628\u0637 \u062d\u062f vmem:<\/strong> \u062e\u0641\u0636 \u062d\u062f \u0627\u0644\u0630\u0627\u0643\u0631\u0629 \u0627\u0644\u0627\u0641\u062a\u0631\u0627\u0636\u064a\u0629 \u0642\u0644\u064a\u0644\u0627\u064b \u062a\u062d\u062a 3GB \u0641\u064a \u0625\u0635\u062f\u0627\u0631 v73 \u0644\u062a\u062c\u0646\u0628 \u0645\u0634\u0627\u0643\u0644 \u0627\u0644\u0630\u0627\u0643\u0631\u0629<\/li>\n
  6. hmx-queue \u0627\u0644\u0645\u0648\u062d\u062f:<\/strong> \u0627\u0633\u062a\u0628\u062f\u0627\u0644 hmx-worker \u0628\u0646\u0638\u0627\u0645 hmx-queue \u064a\u0628\u0633\u0637 \u0627\u0644\u062a\u0646\u0641\u064a\u0630 \u0627\u0644\u0634\u0627\u0645\u0644 \u0648\u064a\u0642\u0644\u0644 \u062f\u0648\u0631\u0627\u062a \u0625\u064a\u0642\u0627\u0638 \u0627\u0644\u062e\u064a\u0648\u0637\u060c (\u0648\u0641\u0642\u0627\u064b \u0644\u0645\u0630\u0643\u0631\u0627\u062a \u0627\u0644\u0625\u0635\u062f\u0627\u0631 \u0627\u0644\u0631\u0633\u0645\u064a\u0629)<\/a><\/li>\n<\/ol>\n

    \u0647\u0630\u0647 \u0627\u0644\u062a\u062d\u0633\u064a\u0646\u0627\u062a \u062a\u0633\u062a\u0647\u062f\u0641 \u0628\u0634\u0643\u0644 \u062e\u0627\u0635 \u0627\u0644\u0645\u0637\u0648\u0631\u064a\u0646 \u0627\u0644\u0630\u064a\u0646 \u064a\u0639\u0645\u0644\u0648\u0646 \u0639\u0644\u0649 \u0623\u062c\u0647\u0632\u0629 \u062a\u062d\u062a\u0648\u064a \u0645\u0639\u0627\u0644\u062c\u0627\u062a Hexagon\u060c \u062e\u0627\u0635\u0629 \u0641\u064a \u0627\u0644\u0628\u064a\u0626\u0627\u062a \u0627\u0644\u0645\u062f\u0645\u062c\u0629 \u0648\u0627\u0644\u0623\u062c\u0647\u0632\u0629 \u0627\u0644\u0645\u062d\u0645\u0648\u0644\u0629. \u0627\u0644\u0641\u0631\u0642 \u0627\u0644\u0623\u0633\u0627\u0633\u064a \u064a\u0643\u0645\u0646 \u0641\u064a \u0643\u064a\u0641\u064a\u0629 \u0625\u062f\u0627\u0631\u0629 \u0639\u0628\u0621 \u0627\u0644\u0639\u0645\u0644 \u0627\u0644\u0645\u062a\u0648\u0627\u0632\u064a – \u0627\u0644\u0646\u0638\u0627\u0645 \u0627\u0644\u062c\u062f\u064a\u062f \u064a\u0633\u0645\u062d \u0628\u062a\u0634\u063a\u064a\u0644 \u0639\u0645\u0644\u064a\u0627\u062a HMX \u0628\u0634\u0643\u0644 \u063a\u064a\u0631 \u0645\u062a\u0632\u0627\u0645\u0646 \u0645\u0639 \u0628\u0627\u0642\u064a \u0627\u0644\u0645\u0639\u0627\u0644\u062c\u0629\u060c \u0645\u0645\u0627 \u064a\u062d\u0633\u0646 \u0627\u0633\u062a\u063a\u0644\u0627\u0644 \u0627\u0644\u0645\u0648\u0627\u0631\u062f \u0628\u0634\u0643\u0644 \u0643\u0628\u064a\u0631.<\/p>\n

    \u0627\u0644\u0625\u0635\u062f\u0627\u0631 \u064a\u0623\u062a\u064a \u0645\u0639 26 \u0628\u0646\u0627\u0621\u064b \u0645\u064f\u062c\u0645\u0651\u0639\u0627\u064b \u0645\u0633\u0628\u0642\u0627\u064b<\/strong> \u064a\u063a\u0637\u064a \u0643\u0644 \u0645\u0646\u0635\u0629 \u062a\u0637\u0648\u064a\u0631 \u0631\u0626\u064a\u0633\u064a\u0629: macOS \u064a\u062d\u0635\u0644 \u0639\u0644\u0649 \u0625\u0635\u062f\u0627\u0631\u064a\u0646 arm64 (\u0639\u0627\u062f\u064a \u0648\u0645\u0639 KleidiAI) \u0648\u0625\u0635\u062f\u0627\u0631 Intel<\/a> x64\u060c Linux \u064a\u0634\u0645\u0644 Ubuntu \u0644\u0644\u0645\u0639\u0645\u0627\u0631\u064a\u0627\u062a x64 \u0648arm64 \u0648s390x \u0645\u0639 \u062f\u0639\u0645 Vulkan \u0648ROCm 7.2 \u0648OpenVINO 2026.0\u060c Windows \u064a\u0642\u062f\u0645 8 \u0645\u062a\u063a\u064a\u0631\u0627\u062a \u0645\u0646 CPU \u0627\u0644\u0623\u0633\u0627\u0633\u064a \u0625\u0644\u0649 CUDA 13.1 \u0648Vulkan \u0648SYCL\u060c \u0628\u064a\u0646\u0645\u0627 openEuler \u064a\u062f\u0639\u0645 \u0645\u0639\u0627\u0644\u062c\u0627\u062a Ascend 310p \u0648910b \u0645\u0639 ACL Graph. \u0644\u0643\u0646 \u0627\u0644\u0645\u0637\u0648\u0631\u064a\u0646 \u0627\u0644\u0630\u064a\u0646 \u0644\u0627 \u064a\u0639\u0645\u0644\u0648\u0646 \u0639\u0644\u0649 Hexagon \u0642\u062f \u0644\u0627 \u064a\u0644\u0627\u062d\u0638\u0648\u0627 \u0641\u0631\u0642\u0627\u064b \u0643\u0628\u064a\u0631\u0627\u064b \u0641\u064a \u0627\u0644\u0623\u062f\u0627\u0621 \u0645\u0642\u0627\u0631\u0646\u0629 \u0628\u0627\u0644\u0625\u0635\u062f\u0627\u0631\u0627\u062a \u0627\u0644\u0633\u0627\u0628\u0642\u0629.<\/p>\n<\/div>\n

    GitHub llama.cpp<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

    llama.cpp b8797 \u062a\u062c\u0644\u0628 6 \u062a\u062d\u0633\u064a\u0646\u0627\u062a \u0647\u0646\u062f\u0633\u064a\u0629 \u0639\u0645\u064a\u0642\u0629 \u0644\u0645\u0639\u0627\u0644\u062c\u0627\u062a Hexagon \u0645\u0639 26 \u0628\u0646\u0627\u0621\u064b \u0645\u064f\u062c\u0645\u0651\u0639\u0627\u064b \u0645\u0633\u0628\u0642\u0627\u064b \u0644\u0644\u0645\u0646\u0635\u0627\u062a \u0627\u0644\u0631\u0626\u064a\u0633\u064a\u0629<\/p>\n","protected":false},"author":1,"featured_media":4236,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[],"class_list":["post-4237","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-learn-use-ai"],"_links":{"self":[{"href":"https:\/\/mojaz.ai\/staging\/wp-json\/wp\/v2\/posts\/4237","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/mojaz.ai\/staging\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/mojaz.ai\/staging\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/mojaz.ai\/staging\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/mojaz.ai\/staging\/wp-json\/wp\/v2\/comments?post=4237"}],"version-history":[{"count":0,"href":"https:\/\/mojaz.ai\/staging\/wp-json\/wp\/v2\/posts\/4237\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/mojaz.ai\/staging\/wp-json\/"}],"wp:attachment":[{"href":"https:\/\/mojaz.ai\/staging\/wp-json\/wp\/v2\/media?parent=4237"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/mojaz.ai\/staging\/wp-json\/wp\/v2\/categories?post=4237"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/mojaz.ai\/staging\/wp-json\/wp\/v2\/tags?post=4237"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}