diff --git a/CMakeLists.txt b/CMakeLists.txt index 98d9feffa..44978da2b 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -190,6 +190,11 @@ if(ONEBIT_JINJA) ${_onebit_jinja_dir}/jinja/caps.cpp # jinja/value.cpp includes the fork's unicode.h, and these are its implementation ${_onebit_jinja_dir}/unicode.cpp) + # engine#2138: from llama.cpp #327 on, jinja/value.cpp works through common_json (common/json.cpp), + # whose JSON_ASSERT is GGML_ASSERT. The binary does not link ggml, so the one function that needs, + # ggml_abort(), comes from app/jinja_ggml_abort.cpp; ggml.h is read for its declaration only. + target_sources(onebit_jinja PRIVATE ${_onebit_jinja_dir}/json.cpp ${CMAKE_SOURCE_DIR}/app/jinja_ggml_abort.cpp) + target_include_directories(onebit_jinja PRIVATE ${CMAKE_SOURCE_DIR}/third_party/llama.cpp/ggml/include) target_include_directories(onebit_jinja PUBLIC ${_onebit_jinja_dir}) target_link_libraries(onebit_jinja PUBLIC nlohmann_json::nlohmann_json) target_compile_definitions(onebit_jinja PUBLIC ONEBIT_JINJA=1) diff --git a/app/jinja_ggml_abort.cpp b/app/jinja_ggml_abort.cpp new file mode 100644 index 000000000..b642659d2 --- /dev/null +++ b/app/jinja_ggml_abort.cpp @@ -0,0 +1,39 @@ +// Copyright 2026 bong-water-water-bong +// SPDX-License-Identifier: Apache-2.0 +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// engine#2138: the one ggml symbol the fork's chat-template sources reach for. +// +// From llama.cpp #327 on, common/jinja/value.cpp works through common_json (common/json.cpp), and +// json.cpp routes nlohmann's JSON_ASSERT to GGML_ASSERT, which calls ggml_abort(). The 1bit binary +// does not link ggml (the HRX and Vulkan servers are separate processes), so onebit_jinja carries +// this definition: the same message ggml prints (file:line: text) on stderr, then abort(). It has +// no abort callback and no backtrace, because nothing in the binary sets one. +#include "ggml.h" + +#include +#include +#include + +extern "C" void ggml_abort(const char* file, int line, const char* fmt, ...) { + std::fflush(stdout); + char message[2048]; + const int offset = std::snprintf(message, sizeof message, "%s:%d: ", file, line); + va_list args; + va_start(args, fmt); + if (offset >= 0 && size_t(offset) < sizeof message) std::vsnprintf(message + offset, sizeof message - size_t(offset), fmt, args); + va_end(args); + std::fprintf(stderr, "%s\n", message); + std::abort(); +} diff --git a/app/npu_template.h b/app/npu_template.h index b6b5548b8..2a448e035 100644 --- a/app/npu_template.h +++ b/app/npu_template.h @@ -61,6 +61,7 @@ #include "jinja/parser.h" #include "jinja/runtime.h" #include "jinja/value.h" +#include "json.h" // common_json, the fork's JSON type global_from_json takes (engine#2138) #endif namespace onebit::npu { @@ -212,7 +213,9 @@ inline std::string render_chat_template(const ChatTemplate& tmpl, const nlohmann if (tmpl.wants_user_system_prompt && !in.contains("user_system_prompt")) in["user_system_prompt"] = ""; // mark_input=true: everything that came in as data (messages, tool arguments, extra request // fields) is marked, so the tokenizer can refuse to parse special tokens inside it. - jinja::global_from_json(ctx, in, /*mark_input=*/true); + // The fork's renderer takes its own common_json (common/json.h, an ordered_json underneath since + // llama.cpp #327), so the round trip through text keeps the order set above. + jinja::global_from_json(ctx, common_json::parse(in.dump()), /*mark_input=*/true); jinja::runtime runtime(ctx); const jinja::value result = runtime.execute(prog); diff --git a/docs/arch-gaps.md b/docs/arch-gaps.md index bd50f00ef..2004d11dd 100644 --- a/docs/arch-gaps.md +++ b/docs/arch-gaps.md @@ -421,6 +421,9 @@ carries `"watch": "report"`, so the census watcher lists new uploads of it as `. does not alert while no pinned backend can serve the class; remove that field (and record `mapped_ok`) when the llama.cpp pin gains `hrm_text`. +* **Mapped since `2c178f44b`** (2026-11-13, the llama.cpp #327 pin engine#2138 moves to): `conversion/hrm_text.py` registers `HrmTextForCausalLM` as `hrm_text`, `src/llama-arch.cpp` registers `LLM_ARCH_HRM_TEXT`, and `src/models/hrm-text.cpp` builds its own graph: a second architecture, not an alias of a supported one. + `registry/significant.json` records the same fact as a `mapped_ok` reason, so `--check-gaps` stays quiet. + ### `dflash2` - `nchapman/Qwen3.5-4B-DFlash2` (also `arc-kumbhar/incoai-Qwen3.8-27B-DFlash2`) Architecture class `DFlash2DraftModel`, no `auto_map` and `library_name: speculators` on the @@ -453,6 +456,9 @@ new uploads of it as `. REVIEWED` and does not alert; remove that field (and rec when a pin bump reaches DFlash2 support (selector + conv heads, per-layer window) and its converter class registers `DFlash2DraftModel`. +* **Mapped since `2c178f44b`** (2026-11-13, the llama.cpp #327 pin engine#2138 moves to): the condition above is met. `conversion/qwen.py` registers `DFlash2DraftModel` with `DFlashDraftModel` and writes `conv_kernel_size`, `conv_group_size`, `selector_rank`, `selector_top_k` and the `candidate_selector` codebooks; `src/models/dflash.cpp` loads the selector tensors and builds the conv and selector heads (`build_dflash2_conv`, `build_dflash2_selector`). The draft no longer drops them. + `registry/significant.json` records the same fact as a `mapped_ok` reason, so `--check-gaps` stays quiet. + ### `tashkeel_nano` - `ahmadsy/tashkeel-nano-1p1m` Architecture class `DiacritizerModel`, `model_type: tashkeel_nano`, no `library_name` and no @@ -550,6 +556,9 @@ the target box today. Reviewed 2026-10-17: **not an alias**. Its `registry/signi entry carries `"watch": "report"`, so the census watcher lists new uploads as `. REVIEWED` and does not alert; remove that field (and record `mapped_ok`) when the pin gains `muse-spark`. +* **Mapped since `2c178f44b`** (2026-11-13, the llama.cpp #327 pin engine#2138 moves to): `conversion/muse_spark.py` registers `MuseSparkForConditionalGeneration` as `muse_spark`, `src/llama-arch.cpp` registers `LLM_ARCH_MUSE_SPARK`, and `src/models/muse-spark.cpp` builds its own graph. + `registry/significant.json` records the same fact as a `mapped_ok` reason, so `--check-gaps` stays quiet. + ### `glm5_next` - `azampatti/GLM-5.3-Flash-AZ` (also `zai-org/GLM-5.3-Flash`) Architecture class `Glm5NextForConditionalGeneration`, `model_type: glm5_next`, no `auto_map` and @@ -602,6 +611,9 @@ counts as supported. Reviewed 2026-10-23: **not an alias**. Its `registry/signif carries `"watch": "report"`, so the census watcher lists new uploads as `. REVIEWED` and does not alert; remove that field (and record `mapped_ok`) when the pin gains `glm5-next`. +* **Mapped since `2c178f44b`** (2026-11-13, the llama.cpp #327 pin engine#2138 moves to): `conversion/glm.py` (text) and `conversion/qwen3vl.py` (the vision tower) register `Glm5NextForConditionalGeneration`, `src/llama-arch.cpp` registers `LLM_ARCH_GLM5_NEXT` with its KDA + MLA cache layout in `src/llama-model.cpp`, and `src/models/glm5-next.cpp` builds its own graph. + `registry/significant.json` records the same fact as a `mapped_ok` reason, so `--check-gaps` stays quiet. + ### 2026-10-27 batch (engine#619) All four classes in that census run are reviewed below. The mirror the run named for @@ -704,6 +716,9 @@ family already ships GGUFs (`ibm-granite/granite-switch-4.1-8b-GGUF`, whose uploads as `. REVIEWED` instead of filing an issue; remove that field (and record `mapped_ok`) when the pin gains `graniteswitch`. +* **Mapped since `2c178f44b`** (2026-11-13, the llama.cpp #327 pin engine#2138 moves to): `conversion/granite.py` registers `GraniteSwitchForCausalLM` as `graniteswitch`, `src/llama-arch.cpp` registers `LLM_ARCH_GRANITE_SWITCH`, and `src/models/granite-switch.cpp` builds its own graph. + `registry/significant.json` records the same fact as a `mapped_ok` reason, so `--check-gaps` stays quiet. + ### 2026-11-05 batch (engine#619) Two classes were reported `UNCOVERED` on 2026-11-05: `Ernie5_5_MoeForCausalLM` (3 models in @@ -816,6 +831,9 @@ inline `config` and the repo's own `config.json`; metadata only, no weights). rotary, 7-tap conv) in the pinned backend, validated against HF reference logits. The family is already 96 models in the census, so this is worth its own issue rather than an alias. +* **Mapped since `2c178f44b`** (2026-11-13, the llama.cpp #327 pin engine#2138 moves to): `conversion/lfm3.py` registers `Lfm3ForCausalLM` as `lfm3`, `src/llama-arch.cpp` registers `LLM_ARCH_LFM3`, and `src/models/lfm3.cpp` builds its own graph: not an alias of `lfm2`/`lfm2moe`. + `registry/significant.json` records the same fact as a `mapped_ok` reason, so `--check-gaps` stays quiet. + #### `needle` - `Amdkn7/needle3` * **Class / model_type**: `NeedleForToolCalling` / `needle`, `library_name: cactus-compute`, diff --git a/registry/architectures.json b/registry/architectures.json index 6a66bbb11..8fab93d86 100644 --- a/registry/architectures.json +++ b/registry/architectures.json @@ -1,13 +1,13 @@ { "about": "HF architecture -> GGUF architecture and the backends whose code accepts it. Generated by tools/registry_build.py from the pinned sources; do not edit.", "sources": { - "llama.cpp (hrx)": "e6c80014eb1c00bd4c1900d18379752227303fbb" + "llama.cpp (hrx)": "ea2d2eb07d5131b2926ee52bbd8cf681a311eb05" }, "counts": { - "hrx": 298, + "hrx": 350, "npu": 9, - "architectures": 298, - "mapped": 298 + "architectures": 350, + "mapped": 350 }, "architectures": { "AfmoeForCausalLM": { @@ -64,12 +64,30 @@ "hrx" ] }, + "BailingMoeV3ForCausalLM": { + "gguf": "bailingmoe3", + "backends": [ + "hrx" + ] + }, + "BailingMoeV3VLForConditionalGeneration": { + "gguf": "bailingmoe3", + "backends": [ + "hrx" + ] + }, "BambaForCausalLM": { "gguf": "granitehybrid", "backends": [ "hrx" ] }, + "BeamForCausalLM": { + "gguf": "beam", + "backends": [ + "hrx" + ] + }, "BertForMaskedLM": { "gguf": "bert", "backends": [ @@ -160,6 +178,12 @@ "hrx" ] }, + "ClefModel": { + "gguf": "clef", + "backends": [ + "hrx" + ] + }, "CodeGenForCausalLM": { "gguf": "codegen", "backends": [ @@ -202,12 +226,42 @@ "hrx" ] }, + "D1Model": { + "gguf": "lfm2", + "backends": [ + "hrx" + ] + }, + "D1OmniModel": { + "gguf": "lfm2", + "backends": [ + "hrx" + ] + }, + "DFlash2DraftModel": { + "gguf": "dflash", + "backends": [ + "hrx" + ] + }, "DFlashDraftModel": { "gguf": "dflash", "backends": [ "hrx" ] }, + "DSparkDraftModel": { + "gguf": "dflash", + "backends": [ + "hrx" + ] + }, + "DSparkSpeculator": { + "gguf": "dflash", + "backends": [ + "hrx" + ] + }, "DbrxForCausalLM": { "gguf": "dbrx", "backends": [ @@ -264,6 +318,12 @@ "hrx" ] }, + "DeepseekV4DSparkModel": { + "gguf": "dflash", + "backends": [ + "hrx" + ] + }, "DeepseekV4ForCausalLM": { "gguf": "deepseek4", "backends": [ @@ -294,6 +354,24 @@ "hrx" ] }, + "Dots3NoteForCausalLM": { + "gguf": "dots3note", + "backends": [ + "hrx" + ] + }, + "Dots3NoteForConditionalGeneration": { + "gguf": "dots3note", + "backends": [ + "hrx" + ] + }, + "Dots3NoteTextForCausalLM": { + "gguf": "dots3note", + "backends": [ + "hrx" + ] + }, "DotsOCRForCausalLM": { "gguf": "qwen2", "backends": [ @@ -324,6 +402,12 @@ "hrx" ] }, + "EmbeddingGemma2Model": { + "gguf": "gemma-embedding2", + "backends": [ + "hrx" + ] + }, "Ernie4_5ForCausalLM": { "gguf": "ernie4_5", "backends": [ @@ -372,6 +456,12 @@ "hrx" ] }, + "ExaoneMoeForCausalLM": { + "gguf": "exaone-moe", + "backends": [ + "hrx" + ] + }, "FalconForCausalLM": { "gguf": "falcon", "backends": [ @@ -541,13 +631,13 @@ ] }, "Gemma4UnifiedForCausalLM": { - "gguf": "gemma4", + "gguf": "gemma-embedding2", "backends": [ "hrx" ] }, "Gemma4UnifiedForConditionalGeneration": { - "gguf": "gemma4", + "gguf": "gemma-embedding2", "backends": [ "hrx" ] @@ -594,6 +684,12 @@ "hrx" ] }, + "Glm5NextForConditionalGeneration": { + "gguf": "glm5-next", + "backends": [ + "hrx" + ] + }, "GlmForCausalLM": { "gguf": "chatglm", "backends": [ @@ -636,12 +732,30 @@ "hrx" ] }, + "GraniteMoeSWAForCausalLM": { + "gguf": "granite_swa", + "backends": [ + "hrx" + ] + }, "GraniteMoeSharedForCausalLM": { "gguf": "granitemoe", "backends": [ "hrx" ] }, + "GraniteSWAForCausalLM": { + "gguf": "granite_swa", + "backends": [ + "hrx" + ] + }, + "GraniteSwitchForCausalLM": { + "gguf": "graniteswitch", + "backends": [ + "hrx" + ] + }, "Grok1ForCausalLM": { "gguf": "grok", "backends": [ @@ -660,12 +774,30 @@ "hrx" ] }, + "H2OLightningModel": { + "gguf": "qwen35", + "backends": [ + "hrx" + ] + }, "HYV3ForCausalLM": { "gguf": "hy_v3", "backends": [ "hrx" ] }, + "HYV4ForCausalLM": { + "gguf": "hy_v4", + "backends": [ + "hrx" + ] + }, + "HrmTextForCausalLM": { + "gguf": "hrm_text", + "backends": [ + "hrx" + ] + }, "HunYuanDenseV1ForCausalLM": { "gguf": "hunyuan-dense", "backends": [ @@ -750,12 +882,24 @@ "hrx" ] }, + "KevModel": { + "gguf": "qwen35", + "backends": [ + "hrx" + ] + }, "KimiK25ForConditionalGeneration": { "gguf": "deepseek2", "backends": [ "hrx" ] }, + "KimiK3ForConditionalGeneration": { + "gguf": "kimi-k3", + "backends": [ + "hrx" + ] + }, "KimiLinearForCausalLM": { "gguf": "kimi-linear", "backends": [ @@ -834,6 +978,12 @@ "hrx" ] }, + "LevModel": { + "gguf": "qwen35", + "backends": [ + "hrx" + ] + }, "Lfm25AudioTokenizer": { "gguf": "lfm2", "backends": [ @@ -852,6 +1002,12 @@ "hrx" ] }, + "Lfm2DSparkDraftModel": { + "gguf": "dflash", + "backends": [ + "hrx" + ] + }, "Lfm2ForCausalLM": { "gguf": "lfm2", "backends": [ @@ -876,6 +1032,42 @@ "hrx" ] }, + "Lfm3D1Model": { + "gguf": "lfm3", + "backends": [ + "hrx" + ] + }, + "Lfm3ForCausalLM": { + "gguf": "lfm3", + "backends": [ + "hrx" + ] + }, + "Lfm3OmniForConditionalGeneration": { + "gguf": "lfm3", + "backends": [ + "hrx" + ] + }, + "Lfm3VlForConditionalGeneration": { + "gguf": "lfm3", + "backends": [ + "hrx" + ] + }, + "LiLiCorrDraftModel": { + "gguf": "dflash", + "backends": [ + "hrx" + ] + }, + "LingDSparkModel": { + "gguf": "dflash", + "backends": [ + "hrx" + ] + }, "LlaMAForCausalLM": { "gguf": "llama", "backends": [ @@ -986,6 +1178,12 @@ "hrx" ] }, + "MapleForCausalLM": { + "gguf": "maple", + "backends": [ + "hrx" + ] + }, "MellumForCausalLM": { "gguf": "mellum", "backends": [ @@ -1018,12 +1216,36 @@ ], "npu_model_type": "minicpm" }, + "MiniCPMOCRForSequenceClassification": { + "gguf": "qwen35", + "backends": [ + "hrx" + ] + }, + "MiniCPMOcrForConditionalGeneration": { + "gguf": "qwen35", + "backends": [ + "hrx" + ] + }, "MiniCPMV4_6ForConditionalGeneration": { "gguf": "qwen35", "backends": [ "hrx" ] }, + "MiniCPMV4_7ForConditionalGeneration": { + "gguf": "qwen35", + "backends": [ + "hrx" + ] + }, + "MiniMaxM1ForCausalLM": { + "gguf": "minimax-01", + "backends": [ + "hrx" + ] + }, "MiniMaxM2ForCausalLM": { "gguf": "minimax-m2", "backends": [ @@ -1042,18 +1264,36 @@ "hrx" ] }, + "MiniMaxText01ForCausalLM": { + "gguf": "minimax-01", + "backends": [ + "hrx" + ] + }, "Ministral3ForCausalLM": { "gguf": "mistral3", "backends": [ "hrx" ] }, + "Ministral3Model": { + "gguf": "mistral3", + "backends": [ + "hrx" + ] + }, "Mistral3ForConditionalGeneration": { "gguf": "mistral3", "backends": [ "hrx" ] }, + "Mistral3Model": { + "gguf": "mistral3", + "backends": [ + "hrx" + ] + }, "MistralForCausalLM": { "gguf": "llama", "backends": [ @@ -1078,6 +1318,12 @@ "hrx" ] }, + "ModernBertDecisionModel": { + "gguf": "modern-bert", + "backends": [ + "hrx" + ] + }, "ModernBertForMaskedLM": { "gguf": "modern-bert", "backends": [ @@ -1102,6 +1348,30 @@ "hrx" ] }, + "MuseGlimmerAssistantModel": { + "gguf": "dflash", + "backends": [ + "hrx" + ] + }, + "MuseGlimmerForConditionalGeneration": { + "gguf": "muse-glimmer", + "backends": [ + "hrx" + ] + }, + "MuseSparkForConditionalGeneration": { + "gguf": "muse-spark", + "backends": [ + "hrx" + ] + }, + "NEOChatModel": { + "gguf": "sensenova_u1", + "backends": [ + "hrx" + ] + }, "NanbeigeForCausalLM": { "gguf": "nanbeige", "backends": [ @@ -1120,6 +1390,12 @@ "hrx" ] }, + "NemotronHPuzzleForCausalLM": { + "gguf": "nemotron_h_moe", + "backends": [ + "hrx" + ] + }, "NeoBERT": { "gguf": "neo-bert", "backends": [ @@ -1138,6 +1414,12 @@ "hrx" ] }, + "NimbleModel": { + "gguf": "qwen35", + "backends": [ + "hrx" + ] + }, "NomicBertModel": { "gguf": "bert", "backends": [ @@ -1198,6 +1480,12 @@ "hrx" ] }, + "OpenJevModel": { + "gguf": "qwen35", + "backends": [ + "hrx" + ] + }, "OrionForCausalLM": { "gguf": "orion", "backends": [ @@ -1300,8 +1588,8 @@ "hrx" ] }, - "QWenForCausalLM": { - "gguf": "qwen", + "PplxDeciderModel": { + "gguf": "qwen35", "backends": [ "hrx" ] @@ -1480,26 +1768,20 @@ "hrx" ] }, - "Qwen4ExpForCausalLM": { - "gguf": "qwen4exp", + "Qwen3_5TextModel": { + "gguf": "qwen35", "backends": [ "hrx" ] }, - "Qwen4ExpForConditionalGeneration": { + "Qwen4ExpForCausalLM": { "gguf": "qwen4exp", "backends": [ "hrx" ] }, - "QwenForCausalLM": { - "gguf": "qwen", - "backends": [ - "hrx" - ] - }, - "QwenLMHeadModel": { - "gguf": "qwen", + "Qwen4ExpForConditionalGeneration": { + "gguf": "qwen4exp", "backends": [ "hrx" ] @@ -1618,6 +1900,18 @@ "hrx" ] }, + "SopranoTextModel": { + "gguf": "qwen3", + "backends": [ + "hrx" + ] + }, + "Spark2_5ForCausalLM": { + "gguf": "spark2_5", + "backends": [ + "hrx" + ] + }, "StableLMEpochForCausalLM": { "gguf": "stablelm", "backends": [ @@ -1726,12 +2020,30 @@ "hrx" ] }, + "VllmD3Model": { + "gguf": "qwen35", + "backends": [ + "hrx" + ] + }, + "VllmDecision2Model": { + "gguf": "vllm-decision2", + "backends": [ + "hrx" + ] + }, "VoxtralForConditionalGeneration": { "gguf": "llama", "backends": [ "hrx" ] }, + "VoxtralRealtimeForConditionalGeneration": { + "gguf": "mistral3", + "backends": [ + "hrx" + ] + }, "WavTokenizerDec": { "gguf": "wavtokenizer-dec", "backends": [ @@ -1827,12 +2139,15 @@ "baichuan", "bailingmoe", "bailingmoe2", + "bailingmoe3", + "beam", "bert", "bitnet", "blackmamba", "bloom", "chameleon", "chatglm", + "clef", "codegen", "codeshell", "cogvlm", @@ -1849,6 +2164,7 @@ "deepseek41", "dflash", "dots1", + "dots3note", "dream", "eagle3", "ernie4_5", @@ -1862,6 +2178,7 @@ "flint", "gemma", "gemma-embedding", + "gemma-embedding2", "gemma2", "gemma3", "gemma3n", @@ -1870,48 +2187,62 @@ "glm-dsa", "glm4", "glm4moe", + "glm5-next", "gpt-oss", "gpt2", "gptj", "gptneo", "gptneox", "granite", + "granite_swa", "granitehybrid", "granitemoe", + "graniteswitch", "grok", "grovemoe", + "hrm_text", "hunyuan-dense", "hunyuan-moe", "hunyuan_vl", "hy_v3", + "hy_v4", "internlm2", "jais", "jais2", "jamba", "jina-bert-v2", "jina-bert-v3", + "k2-horizon", + "kimi-k3", "kimi-linear", + "kolibri1", "laguna", "lfm2", "lfm2moe", + "lfm3", "llada", "llada-moe", "llama", "llama-embed", "llama4", "maincoder", + "maion-coder", "mamba", "mamba2", + "maple", "mellum", "mimo2", "minicpm", "minicpm3", + "minimax-01", "minimax-m2", "minimax-m3", "mistral3", "mistral4", "modern-bert", "mpt", + "muse-glimmer", + "muse-spark", "nanbeige", "nemotron", "nemotron_h", @@ -1934,6 +2265,7 @@ "plamo2", "plamo3", "plm", + "pockettts", "qwen", "qwen2", "qwen2moe", @@ -1943,6 +2275,7 @@ "qwen35moe", "qwen3moe", "qwen3next", + "qwen3tts", "qwen3vl", "qwen3vlmoe", "qwen4exp", @@ -1952,8 +2285,10 @@ "rwkv6qwen2", "rwkv7", "seed_oss", + "sensenova_u1", "smallthinker", "smollm3", + "spark2_5", "stablelm", "starcoder", "starcoder2", @@ -1961,6 +2296,7 @@ "t5", "t5encoder", "talkie", + "vllm-decision2", "wavtokenizer-dec", "xverse", "zamba", diff --git a/registry/significant.json b/registry/significant.json index 9c1a46a63..0d64a1ed5 100644 --- a/registry/significant.json +++ b/registry/significant.json @@ -71,12 +71,14 @@ "HrmTextForCausalLM": { "model_type": "hrm_text", "family": "HRM-Text hierarchical-reasoning LM (Sapient)", - "watch": "report" + "watch": "report", + "mapped_ok": "the pinned backend implements it as its own architecture: third_party/llama.cpp 2c178f44b (llama.cpp #327, engine#2138) maps HrmTextForCausalLM -> hrm_text (conversion/__init__.py, conversion/hrm_text.py), registers LLM_ARCH_HRM_TEXT (src/llama-arch.cpp) and builds its graph in src/models/hrm-text.cpp. docs/arch-gaps.md carries the dated note." }, "DFlash2DraftModel": { "model_type": "dflash2", "family": "DFlash2 speculative drafter (speculators; selector + conv heads the pinned dflash does not run)", - "watch": "report" + "watch": "report", + "mapped_ok": "the pinned backend now runs DFlash2, the condition docs/arch-gaps.md set: third_party/llama.cpp 2c178f44b (llama.cpp #327, engine#2138) registers DFlash2DraftModel with DFlashDraftModel in conversion/qwen.py, which writes conv_kernel_size, conv_group_size, selector_rank, selector_top_k and the candidate_selector codebooks, and src/models/dflash.cpp loads selector_prev/next/hidden and builds them (build_dflash2_conv, build_dflash2_selector). docs/arch-gaps.md carries the dated note." }, "DiacritizerModel": { "model_type": "tashkeel_nano", @@ -101,12 +103,14 @@ "MuseSparkForConditionalGeneration": { "model_type": "muse_spark", "family": "Meta Muse Spark Open (816B-A42B latent-MoE, vision + audio; upstream llama.cpp ba3c05f71, after the pin)", - "watch": "report" + "watch": "report", + "mapped_ok": "the pinned backend implements it as its own architecture: third_party/llama.cpp 2c178f44b (llama.cpp #327, engine#2138) maps MuseSparkForConditionalGeneration -> muse_spark (conversion/__init__.py, conversion/muse_spark.py), registers LLM_ARCH_MUSE_SPARK (src/llama-arch.cpp) and builds its graph in src/models/muse-spark.cpp. docs/arch-gaps.md carries the dated note." }, "Glm5NextForConditionalGeneration": { "model_type": "glm5_next", "family": "GLM-5.3-Flash (KDA linear-attention + DeepSeek sparse attention hybrid, hyper-connections; upstream llama.cpp 649dcb10, after the pin)", - "watch": "report" + "watch": "report", + "mapped_ok": "the pinned backend implements it as its own architecture: third_party/llama.cpp 2c178f44b (llama.cpp #327, engine#2138) maps Glm5NextForConditionalGeneration -> glm5-next (conversion/__init__.py; text in conversion/glm.py, vision tower in conversion/qwen3vl.py), registers LLM_ARCH_GLM5_NEXT (src/llama-arch.cpp, with its KDA + MLA cache layout in src/llama-model.cpp) and builds its graph in src/models/glm5-next.cpp. docs/arch-gaps.md carries the dated note." }, "LiveMemForCausalLM": { "model_type": "livemem", @@ -126,7 +130,8 @@ "GraniteSwitchForCausalLM": { "model_type": "granite_switch", "family": "Granite Switch 4.1 (dense all-attention Granite with N embedded per-token LoRA adapters, a zero adapter slot and weightless switch layers; upstream llama.cpp 157b81fe, not in the pin)", - "watch": "report" + "watch": "report", + "mapped_ok": "the pinned backend implements it as its own architecture: third_party/llama.cpp 2c178f44b (llama.cpp #327, engine#2138) maps GraniteSwitchForCausalLM -> graniteswitch (conversion/__init__.py, conversion/granite.py), registers LLM_ARCH_GRANITE_SWITCH (src/llama-arch.cpp) and builds its graph in src/models/granite-switch.cpp. docs/arch-gaps.md carries the dated note." }, "Ernie5_5_MoeForCausalLM": { "model_type": "ernie5_5_moe", @@ -141,7 +146,8 @@ "Lfm3ForCausalLM": { "model_type": "lfm3", "family": "LiquidAI LFM3 (sandwich norms, sliding-attention layers with partial rotary, 7-tap conv cache, tied embeddings; not an alias of lfm2/lfm2moe, nothing at the pin registers it)", - "watch": "report" + "watch": "report", + "mapped_ok": "the pinned backend implements it as its own architecture: third_party/llama.cpp 2c178f44b (llama.cpp #327, engine#2138) maps Lfm3ForCausalLM -> lfm3 (conversion/__init__.py, conversion/lfm3.py), registers LLM_ARCH_LFM3 (src/llama-arch.cpp) and builds its graph in src/models/lfm3.cpp. docs/arch-gaps.md carries the dated note." }, "NeedleForToolCalling": { "model_type": "needle", diff --git a/tests/census_watch_test.py b/tests/census_watch_test.py index 37773e893..847529a7f 100644 --- a/tests/census_watch_test.py +++ b/tests/census_watch_test.py @@ -348,13 +348,12 @@ def fetch(url): # 15. The classes the 2026-10-14..16 streak flagged (issue #427 and the two un-filed runs # before it) are reviewed, documented and non-alerting. for cls, mid in ( - ("DFlash2DraftModel", "nchapman/Qwen3.5-4B-DFlash2"), + # DFlash2DraftModel used to be here; pin 2c178f44 implements it (asserted as covered, section 15b). ("DiacritizerModel", "ahmadsy/tashkeel-nano-1p1m"), ("MBartForConditionalGeneration", "nhatho1999/orthomni-codeco"), ("BlindForCausalLM", "Vi-LAB/KenT_0.5B_Base"), ("NanochronoForCausalLM", "tensin320/round14_5_114_fitleech_sn38-2023-01"), - # issue #437 (2026-10-17): upstream-only architecture, after the pin. - ("MuseSparkForConditionalGeneration", "Interchained/Muse-Spark-Open-816B-A42B-GGUF"), + # MuseSparkForConditionalGeneration (issue #437) used to be here; pin 2c178f44 implements it (section 15b). # engine #1667 (2026-11-09): the two UNCOVERED classes the daily census raised that day, declined # with their reasons in significant.json. report-only, because nothing a backend can read defines # Liwet's router and nanoGPT's GPT is not GPT-2 under another name. @@ -384,6 +383,22 @@ def fetch(url): check("DeepseekV41ForCausalLM mapped -> no significant alert and no reviewed_report line", res["significant"] == [] and res["reviewed_report"] == []) +# 15b. engine#2138: pin 2c178f44 (llama.cpp #327) implements four more reviewed classes as their own +# architectures (docs/arch-gaps.md, "Mapped since 2c178f44b"). The same uploads that were listed +# REVIEWED are now covered and silent, and significant.json says why (mapped_ok). +MAPPED_2138 = ( + ("DFlash2DraftModel", "nchapman/Qwen3.5-4B-DFlash2"), + ("MuseSparkForConditionalGeneration", "Interchained/Muse-Spark-Open-816B-A42B-GGUF"), + ("GraniteSwitchForCausalLM", "ibm-granite/granite-switch-4.1-8b"), + ("Lfm3ForCausalLM", "D0zy/essay-model-merged"), +) +for cls, mid in MAPPED_2138: + check(f"{cls} is mapped at this pin, with a mapped_ok reason and its dated note", + cls in archs and sig.get(cls, {}).get("mapped_ok") and cls in gap_docs) + rc, res = rc_of([model(mid, cls)], archs, sig) + check(f"{cls} mapped -> covered, exit 0, no significant alert and no reviewed_report line", + rc == 0 and res[cw.COVERED] == [mid] and res["significant"] == [] and res["reviewed_report"] == []) + # --- FastFlowLM org + releases watch (engine#588) ---------------------------------- # # The recorded payloads are what huggingface.co/api/models?author=FastFlowLM and @@ -431,12 +446,14 @@ def flm_state(): # measured there, bench/npu-leaderboard/27b-*.json), so the example moved to a family that is # still HRX-only - gemma3, which FLM also publishes Q4NX packs for. Reading the live registry is # the point: if gemma3 is ever registered for the NPU this check fails and the example moves again. -LFM3_REPO = { - "id": "FastFlowLM/LFM3-1.7B-NPU2", +# A pack for a reviewed-and-declined class. It was FLM's real LFM3-1.7B pack until pin 2c178f44 implemented +# lfm3 (engine#2138); NeedleForToolCalling is still declined, so this listing entry is built for the test. +DECLINED_REPO = { + "id": "FastFlowLM/Needle-Test-NPU2", "lastModified": "2026-10-28T00:00:00.000Z", - "tags": ["base_model:LiquidAI/LFM3-1.7B", "license:other"], + "tags": ["base_model:Amdkn7/needle3", "license:other"], # reviewed, not an alias: registry/significant.json, watch: report -> the pack is listed, not alerted - "config": {"architectures": ["Lfm3ForCausalLM"], "model_type": "lfm3"}, + "config": {"architectures": ["NeedleForToolCalling"], "model_type": "needle"}, "siblings": [{"rfilename": "config.json"}, {"rfilename": "model.q4nx"}], } @@ -473,9 +490,16 @@ def flm_state(): llama = by_id["FastFlowLM/Llama-3.2-1B-NPU2"] check("flm org: a second npu family is reported", llama["npu_model_type"] == "llama") mini = by_id["FastFlowLM/MiniCPM-V-4.7-1B-NPU2"] -check("flm org: an unmapped architecture is unmapped, not a guess", +# engine#2138: from pin 2c178f44 the fork converts MiniCPM-V 4.7 (a qwen35 text model plus its own +# projector), so the live registry maps it for hrx only and its pack reads as an NPU gap. The unmapped +# case is the same listing against a registry without the class - the one before that pin. +pre_2138_archs = {k: v for k, v in live_archs.items() if k != "MiniCPMV4_7ForConditionalGeneration"} +check("flm org: a family mapped for hrx only reads as an NPU gap, not as served", mini["architecture"] == "MiniCPMV4_7ForConditionalGeneration" - and not mini["mapped"] and cw.unmapped(mini) and not mini["npu"]) + and mini["mapped"] and not mini["npu"] and cw.unmapped(mini)) +mini_pre = {r["id"]: r for r in cw.sweep_org(pre_2138_archs, fetch=flm_fetch)}["FastFlowLM/MiniCPM-V-4.7-1B-NPU2"] +check("flm org: an unmapped architecture is unmapped, not a guess", + not mini_pre["mapped"] and cw.unmapped(mini_pre) and not mini_pre["npu"]) gguf = by_id["FastFlowLM/Qwen3.5-0.8B-gguf"] check("flm org: a config-less gguf drop is a derivative, not a new family", gguf["derivative"] and not cw.unmapped(gguf) and not gguf["mapped"]) @@ -511,14 +535,15 @@ def counting_detail(url): check("flm sizes: a failed detail fetch keeps the repo (size 0, no crash)", [f["size"] for f in kept["q4nx"]] == [0]) -# engine#1451: `Lfm3ForCausalLM` and `DFlash2DraftModel` are recorded in registry/significant.json as +# engine#1451: `Lfm3ForCausalLM` and `DFlash2DraftModel` were recorded in registry/significant.json as # reviewed-not-an-alias. The class census already printed them `. REVIEWED`; the FLM half printed # `!! NEW FLM ... registry: none` forever, which is what kept a daily issue open on an answered question. -_declined_rows = cw.sweep_org(live_archs, fetch=lambda url: fixture("org.json") + [LFM3_REPO]) +# Both are mapped from pin 2c178f44 (engine#2138), so a still-declined class stands in for them. +_declined_rows = cw.sweep_org(live_archs, fetch=lambda url: fixture("org.json") + [DECLINED_REPO]) _buf = io.StringIO() with contextlib.redirect_stdout(_buf): cw.report_flm(_declined_rows, cw.sweep_releases(fetch=flm_fetch), - [r for r in _declined_rows if r["id"] == "FastFlowLM/LFM3-1.7B-NPU2"], [], answers=NO_ANSWERS) + [r for r in _declined_rows if r["id"] == DECLINED_REPO["id"]], [], answers=NO_ANSWERS) _declined_listing = _buf.getvalue() # 18. One line per repo, `!! NEW FLM` only for a family we cannot serve, and the exit code @@ -533,14 +558,19 @@ def counting_detail(url): check("flm report: one line per repo", len([ln for ln in first.splitlines() if ln.startswith(FLM_LINE_PREFIXES)]) == 5) -check("flm report: the unknown family is marked !! NEW FLM", - " !! NEW FLM FastFlowLM/MiniCPM-V-4.7-1B-NPU2" in first) +check("flm report: a family mapped for hrx only is marked !! NPU FLM (engine#2138: MiniCPM-V 4.7)", + " !! NPU FLM FastFlowLM/MiniCPM-V-4.7-1B-NPU2" in first) +_pre_buf = io.StringIO() +with contextlib.redirect_stdout(_pre_buf), contextlib.redirect_stderr(io.StringIO()): + _pre_rc, _ = cw.watch_flm(pre_2138_archs, flm_state(), fetch=flm_fetch, detail=flm_detail, answers=NO_ANSWERS) +check("flm report: the unknown family is marked !! NEW FLM (the registry before pin 2c178f44)", + " !! NEW FLM FastFlowLM/MiniCPM-V-4.7-1B-NPU2" in _pre_buf.getvalue() and _pre_rc == 1) check("flm report: a reviewed-and-declined class is listed, not alerted (engine#1451)", - " . FLM FastFlowLM/LFM3-1.7B-NPU2" in _declined_listing - and "!! NEW FLM FastFlowLM/LFM3-1.7B-NPU2" not in _declined_listing + " . FLM " + DECLINED_REPO["id"] in _declined_listing + and "!! NEW FLM " + DECLINED_REPO["id"] not in _declined_listing and "registry: none (reviewed, not an alias" in _declined_listing) check("flm report: a reviewed-and-declined class is not counted as an unmapped family", - "LFM3-1.7B-NPU2" not in _declined_listing.split("publish a family this engine does not serve")[-1] + DECLINED_REPO["id"] not in _declined_listing.split("publish a family this engine does not serve")[-1] if "publish a family this engine does not serve" in _declined_listing else True) check("flm report: a mapped family is not marked", " + FLM FastFlowLM/Qwen3-8B-NPU2" in first and "!! NEW FLM FastFlowLM/Qwen3-8B-NPU2" not in first) @@ -749,7 +779,7 @@ def fail_fetch(url): "LiveMemForCausalLM": ("livemem", "Popbones/LiveMem-RL-mlx-q4-backbone"), "CDMMoELMPoincare": ("cdm_moe_poincare", "DuoNeural/Lattice-7-Hermes-CDM-Poincare-Instruct"), "Senryu2ForCausalLM": ("senryu2", "KerzgrAI/Senryu-2.2-1B"), - "GraniteSwitchForCausalLM": ("granite_switch", "ibm-granite/granite-switch-4.1-8b"), + # GraniteSwitchForCausalLM was the fourth; pin 2c178f44 implements it (section 15b, engine#2138). } for cls, (model_type, mid) in batch_619.items(): entry = sig.get(cls) or {} @@ -761,8 +791,8 @@ def fail_fetch(url): rc == 0 and res["reviewed_report"] == [cls] and res["significant"] == [] and res[cw.UNCOVERED] == [mid]) _, out = text_of([model("ibm-granite/granite-switch-4.1-8b", "GraniteSwitchForCausalLM")], archs, sig) -check("granite switch is printed as REVIEWED, not UNCOVERED", - ". REVIEWED GraniteSwitchForCausalLM" in out and "UNCOVERED" not in out) +check("granite switch is not printed as REVIEWED or UNCOVERED now that the pin implements it (engine#2138)", + "REVIEWED GraniteSwitchForCausalLM" not in out and "UNCOVERED" not in out) variant_619 = { "LNewqForConditionalGeneration": "crussella0129/LNewq", @@ -784,7 +814,7 @@ def fail_fetch(url): with open(os.path.join(ROOT, "registry", "census.json")) as f: counts = json.load(f)["raw"]["counts"] check("census.json records every engine#619 class", - all(counts.get(cls) for cls in list(batch_619) + list(variant_619))) + all(counts.get(cls) for cls in list(batch_619) + list(variant_619) + ["GraniteSwitchForCausalLM"])) check("census.json shows granite switch is not a one-off", counts["GraniteSwitchForCausalLM"] > 1) @@ -824,13 +854,15 @@ def fail_fetch(url): check(f"{repo_id} is recorded in docs/npu.md", repo_id in npu_docs) if not cls: continue - if cls == "HYV3ForCausalLM": - check("Hy-Memory's hy_v3 is mapped for hrx only, so its pack is an NPU gap", + if cls in archs: + # Hy-Memory's hy_v3 from the start; from pin 2c178f44 (engine#2138) also MiniCPM-V 4.7, MiniCPM-OCR, + # DFlash2 and LFM3, which the fork now converts. Mapped for hrx only, so the pack is an NPU gap. + check(f"{cls} is mapped for hrx only, so {repo_id} is an NPU gap", archs[cls]["backends"] == ["hrx"] and "npu" not in archs[cls]["backends"] and cls in npu_docs) hy = cw.flm_repo({"id": repo_id, "config": {"architectures": [cls]}, "siblings": [{"rfilename": "model.q4nx"}]}, archs) - check("Hy-Memory pack resolves to the recorded npu gap, not a mapping", + check(f"{repo_id} resolves to the recorded npu gap, not a mapping", hy["mapped"] and cw.npu_gap(hy) and cw.unmapped(hy) and not hy["npu"]) else: check(f"{cls} stays out of the registry (no GGUF alias claimed)", @@ -845,7 +877,7 @@ def fail_fetch(url): # report-only, and the FastFlowLM NPU rows are answered family by family in docs/npu.md # instead of being mapped as GGUF aliases or given an `npu` backend they do not have. batch_1106 = { - "Lfm3ForCausalLM": ("lfm3", "D0zy/essay-model-merged"), + # Lfm3ForCausalLM was the other; pin 2c178f44 implements it (section 15b, engine#2138). "NeedleForToolCalling": ("needle", "Amdkn7/needle3"), } for cls, (model_type, mid) in batch_1106.items(): diff --git a/third_party/llama.cpp b/third_party/llama.cpp index e6c80014e..ea2d2eb07 160000 --- a/third_party/llama.cpp +++ b/third_party/llama.cpp @@ -1 +1 @@ -Subproject commit e6c80014eb1c00bd4c1900d18379752227303fbb +Subproject commit ea2d2eb07d5131b2926ee52bbd8cf681a311eb05 diff --git a/tools/ci/hrx_f16_accumulators.txt b/tools/ci/hrx_f16_accumulators.txt index 040b637ab..34a234728 100644 --- a/tools/ci/hrx_f16_accumulators.txt +++ b/tools/ci/hrx_f16_accumulators.txt @@ -25,6 +25,19 @@ # regenerates this file and tools/ci/hrx_f16_narrowing.txt; a new entry is drift the ratchet # refuses, and a line whose file no longer has a mul-updated f16 fragment is stale and should be # dropped as the corpus is fixed. +# engine#2138 (pin ca5aac715): 11 accumulators added, all from AMD's composable WMMA kernels in #327: the f16 fragment +# of the matmul / matmul_id tile motifs and flash_attention/{decode_partials,pv_block_wmma}. Nothing removed. +loom-libs/motifs/flash_attention/decode_partials.loom 2 +loom-libs/motifs/flash_attention/pv_block_wmma.loom 2 +loom-libs/motifs/matmul/tiled_input_f32_codebook_dequant_accumulate_f32.loom 4 +loom-libs/motifs/matmul/tiled_input_f32_dequant_accumulate_f32.loom 4 +loom-libs/motifs/matmul/tiled_input_f32_dequant_accumulate_f32_token64.loom 4 +loom-libs/motifs/matmul/tiled_pair_input_f32_dequant_accumulate_f32.loom 4 +loom-libs/motifs/matmul/vector_q6_f32_body.loom 1 +loom-libs/motifs/matmul_id/f16_down_schedule.loom 4 +loom-libs/motifs/matmul_id/pair_input_f32_dequant_accumulate_f32.loom 4 +loom-libs/motifs/matmul_id/skinny_input_f32_dequant_accumulate_f32.loom 4 +loom-libs/motifs/matmul_id/tiled_input_f32_dequant_accumulate_f32.loom 4 loom-libs/motifs/mul_mat_id_f16_f16_wmma_core.loom 4 loom-libs/motifs/mul_mat_id_swiglu_f16_f16_wmma_body.loom 2 loom-libs/ops/flash_attention_decode_split_f32_f16_wmma.loom 2 diff --git a/tools/ci/hrx_f16_guards.txt b/tools/ci/hrx_f16_guards.txt index ac6e46579..a1353e16b 100644 --- a/tools/ci/hrx_f16_guards.txt +++ b/tools/ci/hrx_f16_guards.txt @@ -25,6 +25,7 @@ # regenerates this file and tools/ci/hrx_f16_narrowing.txt; a new entry is drift the ratchet # refuses, and a line whose file no longer has a mul-updated f16 fragment is stale and should be # dropped as the corpus is fixed. +loom-libs/motifs/flash_attention/online_softmax.loom loom-libs/motifs/mul_mat_f32_f32_wmma_core.loom loom-libs/motifs/mul_mat_id_f32_f32_wmma_core.loom loom-libs/motifs/mul_mat_id_q4k_f16_wmma_projection.loom diff --git a/tools/ci/hrx_f16_narrowing.txt b/tools/ci/hrx_f16_narrowing.txt index 8b7441ec5..a3b263ab7 100644 --- a/tools/ci/hrx_f16_narrowing.txt +++ b/tools/ci/hrx_f16_narrowing.txt @@ -19,12 +19,63 @@ # perfectly correct. What the list buys is that a NEW narrowing body cannot appear without someone # writing it down and saying what feeds it -- and that is the open question on engine#1962: which of # these are protected by a producer-side prescale cannot be answered by reading them. +# +# engine#2138 (pin 2f7f04f0d..ca5aac715, llama.cpp #327's AMD merge + #345): 34 bodies added, 1 dropped, 1 recount. +# What feeds each new body is the producer of its fptrunc operand. A script classified them; I checked the +# activation and projection classes against matmul_id/skinny_input_f32_dequant_accumulate_f32, +# matmul/tiled_input_f32_dequant_accumulate_f32 and matmul_id/f16_projection_q4k (the rest share their shape): +# f32 ACTIVATION narrowed for an f16 WMMA tile, no guard marker (engine#1962's open question; nothing upstream of +# these is known to prescale): matmul/tiled_input_f32_{,codebook_}dequant_accumulate_f32{,_token64}, +# matmul/tiled_pair_input_f32_dequant_accumulate_f32, matmul/vector_q6_f32_body, matmul_id/{skinny,tiled,pair}_input_ +# f32_dequant_accumulate_f32, matmul_id/f16_pair_accumulate, copy/convert_f32_f16. +# ROUTED f16 projections, AMD's composable copies (matmul_id/f16_projection_q{4,5,6}k): the dequantized value and the +# loaded activation are narrowed with NO guard marker, where this line's mul_mat_id_q{4,5,6}k_f16_wmma_projection carry +# one (hrx_f16_guards.txt, engine#2127). They are library sources of the dispatched +# ggml_mul_mat_id_swiglu_f32_f32_wmma{,_iq3s_table}. Whether those kernels reach these functions is not established. +# DEQUANTIZED WEIGHTS (bounded by the format's scale): the dequant callers in the matmul motifs above, the +# rmsnorm-weight multiplies in gated_delta_net/rmsnorm_gate and ops/rows/get_rows_rmsnorm_binary_q8_1_x4_f16; +# dequant.loom 20 -> 21 is #345's ggml_codebook_f16_vector4. +# ATTENTION probabilities and softmax scales (in [0,1] or a 1/sum): flash_attention/{online_softmax,decode_partials, +# qk_block_wmma,pv_block_wmma,publish_f32}. +# PUBLISHED OUTPUTS stored as f16 (a result's own range; a KV cache or a normalized row): rows/cache_publish, +# attention/rmsnorm_neox_cache_f16, ops/llm/llm_attention_{k_matmul_rope,v_matmul}_set_rows_vector_f32_f32, +# matmul/{tiled,vector}_publish, matmul_id/f16_pair_binary_publish, normalization/{publish_f16,rmsnorm_binary}. +# q8_1 BLOCK scale d and scaled sum s (engine#2186 item 2's class): quantize_q8_1_x4_group, normalization/ +# rmsnorm_q8_1_x4, matmul_id/vector_pair. +# STATE: gated_delta_net/scan_state_head128 (54 lines, the recurrent state and its deltas). +# TEST-ONLY kernel: ops/flash_attention/decode_split_next_q8_test. +# Dropped: qwen_moe/qwen3_moe/attention_postprocess_f32_f16.loom no longer narrows (3 -> 0 at this pin). +loom-libs/motifs/attention/rmsnorm_neox_cache_f16.loom 3 loom-libs/motifs/bf16_f16.loom 1 -loom-libs/motifs/dequant.loom 20 +loom-libs/motifs/copy/convert_f32_f16.loom 2 +loom-libs/motifs/dequant.loom 21 loom-libs/motifs/dequant_1bit.loom 5 loom-libs/motifs/dequant_prism.loom 2 loom-libs/motifs/f32_f16.loom 1 +loom-libs/motifs/flash_attention/decode_partials.loom 2 +loom-libs/motifs/flash_attention/online_softmax.loom 1 +loom-libs/motifs/flash_attention/publish_f32.loom 8 +loom-libs/motifs/flash_attention/pv_block_wmma.loom 2 +loom-libs/motifs/flash_attention/qk_block_wmma.loom 1 +loom-libs/motifs/gated_delta_net/rmsnorm_gate.loom 1 +loom-libs/motifs/gated_delta_net/scan_state_head128.loom 54 loom-libs/motifs/llm_attention_qkv_matmul_postops.loom 1 +loom-libs/motifs/matmul/tiled_input_f32_codebook_dequant_accumulate_f32.loom 2 +loom-libs/motifs/matmul/tiled_input_f32_dequant_accumulate_f32.loom 3 +loom-libs/motifs/matmul/tiled_input_f32_dequant_accumulate_f32_token64.loom 1 +loom-libs/motifs/matmul/tiled_pair_input_f32_dequant_accumulate_f32.loom 5 +loom-libs/motifs/matmul/tiled_publish.loom 1 +loom-libs/motifs/matmul/vector_publish.loom 1 +loom-libs/motifs/matmul/vector_q6_f32_body.loom 12 +loom-libs/motifs/matmul_id/f16_pair_accumulate.loom 1 +loom-libs/motifs/matmul_id/f16_pair_binary_publish.loom 1 +loom-libs/motifs/matmul_id/f16_projection_q4k.loom 5 +loom-libs/motifs/matmul_id/f16_projection_q5k.loom 5 +loom-libs/motifs/matmul_id/f16_projection_q6k.loom 2 +loom-libs/motifs/matmul_id/pair_input_f32_dequant_accumulate_f32.loom 1 +loom-libs/motifs/matmul_id/skinny_input_f32_dequant_accumulate_f32.loom 1 +loom-libs/motifs/matmul_id/tiled_input_f32_dequant_accumulate_f32.loom 1 +loom-libs/motifs/matmul_id/vector_pair.loom 2 loom-libs/motifs/mul_mat_f32_f32_wmma_core.loom 3 loom-libs/motifs/mul_mat_id_f16_f16_wmma_core.loom 5 loom-libs/motifs/mul_mat_id_f32_f32_wmma_core.loom 1 @@ -34,16 +85,24 @@ loom-libs/motifs/mul_mat_id_q6k_f16_wmma_projection.loom 2 loom-libs/motifs/mul_mat_id_swiglu_f16_f16_accumulate.loom 1 loom-libs/motifs/mul_mat_id_swiglu_f16_f16_wmma_body.loom 1 loom-libs/motifs/mul_mat_quantized_f16_prefill.loom 4 +loom-libs/motifs/normalization/publish_f16.loom 2 +loom-libs/motifs/normalization/rmsnorm_binary.loom 2 +loom-libs/motifs/normalization/rmsnorm_q8_1_x4.loom 2 loom-libs/motifs/publish_f32.loom 1 loom-libs/motifs/q4_k_f16.loom 4 loom-libs/motifs/q6_k_f16.loom 5 loom-libs/motifs/q8_0_f16.loom 1 loom-libs/motifs/q8_1_f16.loom 1 loom-libs/motifs/quantize_q8_1_x4.loom 2 +loom-libs/motifs/quantize_q8_1_x4_group.loom 2 +loom-libs/motifs/rows/cache_publish.loom 2 loom-libs/ops/copy_f32.loom 2 +loom-libs/ops/flash_attention/decode_split_next_q8_test.loom 3 loom-libs/ops/flash_attention_decode_split_f32_f16_wmma.loom 2 loom-libs/ops/flash_attention_f32_f16_wmma.loom 9 loom-libs/ops/gated_delta_net_f32_wmma.loom 56 +loom-libs/ops/llm/llm_attention_k_matmul_rope_set_rows_vector_f32_f32.loom 1 +loom-libs/ops/llm/llm_attention_v_matmul_set_rows_vector_f32_f32.loom 2 loom-libs/ops/llm_attention_k_matmul_rope_set_rows_decode_f32_f32.loom 1 loom-libs/ops/llm_attention_v_matmul_set_rows_decode_f32_f32.loom 2 loom-libs/ops/mul_mat_id_swiglu_f32_f32_wmma.loom 1 @@ -54,11 +113,11 @@ loom-libs/ops/mul_mat_symmetric_i4_wmma.loom 1 loom-libs/ops/mul_mat_symmetric_i8_wmma.loom 1 loom-libs/ops/rmsnorm_binary_f32.loom 2 loom-libs/ops/rope_set_rows_f32.loom 2 +loom-libs/ops/rows/get_rows_rmsnorm_binary_q8_1_x4_f16.loom 1 loom-libs/ops/set_rows.loom 2 loom-libs/ops/small_rows_f32.loom 1 qwen_moe/ggml/linear_q6k_q8_1_x4.loom 1 qwen_moe/ggml/quantize_q8_1_x4.loom 2 -qwen_moe/qwen3_moe/attention_postprocess_f32_f16.loom 3 qwen_moe/qwen3_moe/attention_prepare_quantized.loom 2 qwen_moe/qwen3_moe/dense_linear_quantized_f16_wmma.loom 5 qwen_moe/qwen3_moe/flash_attention_decode_f32_f16_wmma.loom 11