From 0aa3f0e12f77469d3997e60e2205db970bf1f68f Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Tue, 13 Oct 2026 06:21:35 -0700 Subject: [PATCH 01/11] Pin unslothai#264, #265, #266: MoE prompt processing, Q2_0 x86 SIMD and mmap populate --- scripts/unsloth/feature-checks.json | 11 ++++++++++- scripts/unsloth/pr-set.json | 5 ++++- 2 files changed, 14 insertions(+), 2 deletions(-) diff --git a/scripts/unsloth/feature-checks.json b/scripts/unsloth/feature-checks.json index 7c4e387b98af..bf20692b33c4 100644 --- a/scripts/unsloth/feature-checks.json +++ b/scripts/unsloth/feature-checks.json @@ -86,6 +86,13 @@ { "kind": "backend-op", "op": "TOPK_QSA" } ] }, + "cuda-moe-prompt": { + "owner": "unslothai#264", + "checks": [ + { "kind": "backend-op", "op": "MUL_MAT_ID", "params": "^(?!.*dup=1)" }, + { "kind": "backend-op", "op": "GET_ROWS" } + ] + }, "projector-registry": { "owner": "unslothai#176", "checks": [ @@ -102,6 +109,8 @@ "unslothai#241": "CUDA graph cache keyed by shape with no count cap (tensor split keeps replaying its 2 * n_layers + 1 graphs per device), macOS row prefetch and the F16 glm5-next indexer cache under a quantized -ctk (the NextN draft graph is upstream since ggml-org#29928); the first needs a CUDA host (multi-GPU for the cap), the second has no observable output difference, the last needs a real GLM-5-Next model", "unslothai#243": "glm5-next compute buffer size; a memory-layout change with no observable output difference, and the size only shows on a CUDA host with a GLM-5-Next model", "unslothai#251": "--moe-cache-mib auto sizes the ggml-org#29887 MoE expert cache from the fit; it only acts when a MoE model spills on a single GPU, the flag is off by default, and no runner in the pipeline has a GPU", - "unslothai#270": "two-subgroup coopmat1 FA tile, chosen only where the four-subgroup tile does not fit shared memory (32 KiB, e.g. the AMD proprietary driver); needs such a Vulkan device, and no runner in the pipeline has one" + "unslothai#270": "two-subgroup coopmat1 FA tile, chosen only where the four-subgroup tile does not fit shared memory (32 KiB, e.g. the AMD proprietary driver); needs such a Vulkan device, and no runner in the pipeline has one", + "unslothai#265": "x86 SIMD vec_dot for Q2_0; test-backend-ops compares a backend against the CPU reference, so a CPU kernel cannot be checked against itself (the PR compares SIMD builds against the scalar build as a second CPU backend instead)", + "unslothai#266": "MADV_POPULATE_READ on the non-lazy ranges when MAP_POPULATE is skipped; a page-fault and load-time change with no observable output difference, so there is no feature surface to probe" } } diff --git a/scripts/unsloth/pr-set.json b/scripts/unsloth/pr-set.json index 2eebe6a5db27..6d176015aa6c 100644 --- a/scripts/unsloth/pr-set.json +++ b/scripts/unsloth/pr-set.json @@ -33,6 +33,9 @@ "https://github.com/unslothai/llama.cpp/pull/241/commits/20dd977eae5e1a3b0633bba85f03d8b09692379f", "https://github.com/unslothai/llama.cpp/pull/243/commits/f67cec2bd6806d5294e2abc5d62a16405dd3eb74", "https://github.com/unslothai/llama.cpp/pull/251/commits/07208276ed5ed4bfc27a152bbbc8c030a09f870a", - "https://github.com/unslothai/llama.cpp/pull/270/commits/7968eda613e427ad5ccbab61df43246a9a82c53a" + "https://github.com/unslothai/llama.cpp/pull/270/commits/7968eda613e427ad5ccbab61df43246a9a82c53a", + "https://github.com/unslothai/llama.cpp/pull/264/commits/456e23cd364608c4e565f5bde375ecfaf6f3ec9d", + "https://github.com/unslothai/llama.cpp/pull/265/commits/3f9cad27767b9c38f9461ba4d5515e2433bbb278", + "https://github.com/unslothai/llama.cpp/pull/266/commits/88035b0f74d0fd076df2012a957950fee0c4615f" ] } From 1c32f6e6c48a46ae8bae56ed0800b62c2a877d6c Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Wed, 14 Oct 2026 09:56:18 -0700 Subject: [PATCH 02/11] pr-set: drop unslothai#265, upstream ggml-org#30447 adds the same AVX2 Q2_0 dot (in b11612) --- scripts/unsloth/feature-checks.json | 1 - scripts/unsloth/pr-set.json | 1 - 2 files changed, 2 deletions(-) diff --git a/scripts/unsloth/feature-checks.json b/scripts/unsloth/feature-checks.json index 22886beb876c..8696174205a5 100644 --- a/scripts/unsloth/feature-checks.json +++ b/scripts/unsloth/feature-checks.json @@ -109,7 +109,6 @@ "unslothai#243": "glm5-next compute buffer size; a memory-layout change with no observable output difference, and the size only shows on a CUDA host with a GLM-5-Next model", "unslothai#251": "--moe-cache-mib auto sizes the ggml-org#29887 MoE expert cache from the fit; it only acts when a MoE model spills on a single GPU, the flag is off by default, and no runner in the pipeline has a GPU", "unslothai#270": "two-subgroup coopmat1 FA tile, chosen only where the four-subgroup tile does not fit shared memory (32 KiB, e.g. the AMD proprietary driver); needs such a Vulkan device, and no runner in the pipeline has one", - "unslothai#265": "x86 SIMD vec_dot for Q2_0; test-backend-ops compares a backend against the CPU reference, so a CPU kernel cannot be checked against itself (the PR compares SIMD builds against the scalar build as a second CPU backend instead)", "unslothai#266": "MADV_POPULATE_READ on the non-lazy ranges when MAP_POPULATE is skipped; a page-fault and load-time change with no observable output difference, so there is no feature surface to probe" } } diff --git a/scripts/unsloth/pr-set.json b/scripts/unsloth/pr-set.json index af874757141e..f16804d97881 100644 --- a/scripts/unsloth/pr-set.json +++ b/scripts/unsloth/pr-set.json @@ -34,7 +34,6 @@ "https://github.com/unslothai/llama.cpp/pull/251/commits/07208276ed5ed4bfc27a152bbbc8c030a09f870a", "https://github.com/unslothai/llama.cpp/pull/270/commits/7968eda613e427ad5ccbab61df43246a9a82c53a", "https://github.com/unslothai/llama.cpp/pull/264/commits/456e23cd364608c4e565f5bde375ecfaf6f3ec9d", - "https://github.com/unslothai/llama.cpp/pull/265/commits/3f9cad27767b9c38f9461ba4d5515e2433bbb278", "https://github.com/unslothai/llama.cpp/pull/266/commits/88035b0f74d0fd076df2012a957950fee0c4615f" ] } From c3cb48813a2d98b98400935833a698d7ea99a70d Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Wed, 14 Oct 2026 11:12:49 -0700 Subject: [PATCH 03/11] pr-set: repin unslothai#243 onto b11617 (upstream #30480 conflict), #251 and #264 review fixes --- scripts/unsloth/pr-set.json | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/scripts/unsloth/pr-set.json b/scripts/unsloth/pr-set.json index f16804d97881..3a4e77e78690 100644 --- a/scripts/unsloth/pr-set.json +++ b/scripts/unsloth/pr-set.json @@ -30,10 +30,10 @@ "https://github.com/unslothai/llama.cpp/pull/239/commits/bec216444552a20bc39bf2156fa5396154970608", "https://github.com/unslothai/llama.cpp/pull/240/commits/d9dd1810da0075ff0000a352055ab246b2738d5e", "https://github.com/unslothai/llama.cpp/pull/241/commits/20dd977eae5e1a3b0633bba85f03d8b09692379f", - "https://github.com/unslothai/llama.cpp/pull/243/commits/f67cec2bd6806d5294e2abc5d62a16405dd3eb74", - "https://github.com/unslothai/llama.cpp/pull/251/commits/07208276ed5ed4bfc27a152bbbc8c030a09f870a", + "https://github.com/unslothai/llama.cpp/pull/243/commits/f427abb1c8b0cc66c138fa1af42c9eaa368eec82", + "https://github.com/unslothai/llama.cpp/pull/251/commits/2f5ad7f7908d2bd655a82faf2f7b5b9a984f787f", "https://github.com/unslothai/llama.cpp/pull/270/commits/7968eda613e427ad5ccbab61df43246a9a82c53a", - "https://github.com/unslothai/llama.cpp/pull/264/commits/456e23cd364608c4e565f5bde375ecfaf6f3ec9d", + "https://github.com/unslothai/llama.cpp/pull/264/commits/5bb68456bc34238cedb50c8683495c6d6f924177", "https://github.com/unslothai/llama.cpp/pull/266/commits/88035b0f74d0fd076df2012a957950fee0c4615f" ] } From f39249cb4d69dcffa606e621f7bb7636a024d414 Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Wed, 14 Oct 2026 12:28:22 -0700 Subject: [PATCH 04/11] pr-set: repin unslothai#149 (warn only after a successful managed allocation) --- scripts/unsloth/pr-set.json | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/scripts/unsloth/pr-set.json b/scripts/unsloth/pr-set.json index 3a4e77e78690..30b6a4ca35b5 100644 --- a/scripts/unsloth/pr-set.json +++ b/scripts/unsloth/pr-set.json @@ -25,7 +25,7 @@ "https://github.com/unslothai/llama.cpp/pull/95/commits/3db8cb5b2e9bf291057b9f19960e8601a162da81", "https://github.com/unslothai/llama.cpp/pull/158/commits/fb9b9462041ad58cde9a274e7984c3c6fb01b279", "https://github.com/unslothai/llama.cpp/pull/157/commits/6c6da89266ba7839d825c9997782af4f4d26b81b", - "https://github.com/unslothai/llama.cpp/pull/149/commits/30d39d74fd9cad3224c30bdda3a9def6c9aa6719", + "https://github.com/unslothai/llama.cpp/pull/149/commits/efa8b0e689ac8ba13f5840d9c26aa227ffd07153", "https://github.com/unslothai/llama.cpp/pull/176/commits/e2519ebc09468b477847a7529834a8da998ed2da", "https://github.com/unslothai/llama.cpp/pull/239/commits/bec216444552a20bc39bf2156fa5396154970608", "https://github.com/unslothai/llama.cpp/pull/240/commits/d9dd1810da0075ff0000a352055ab246b2738d5e", From 5fabca528462823c9ba1a4a6f230b0c3109e0106 Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Wed, 14 Oct 2026 12:35:10 -0700 Subject: [PATCH 05/11] pr-set: repin unslothai#251 (keep the auto cache size across a reload) --- scripts/unsloth/pr-set.json | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/scripts/unsloth/pr-set.json b/scripts/unsloth/pr-set.json index 30b6a4ca35b5..92f573171f57 100644 --- a/scripts/unsloth/pr-set.json +++ b/scripts/unsloth/pr-set.json @@ -31,7 +31,7 @@ "https://github.com/unslothai/llama.cpp/pull/240/commits/d9dd1810da0075ff0000a352055ab246b2738d5e", "https://github.com/unslothai/llama.cpp/pull/241/commits/20dd977eae5e1a3b0633bba85f03d8b09692379f", "https://github.com/unslothai/llama.cpp/pull/243/commits/f427abb1c8b0cc66c138fa1af42c9eaa368eec82", - "https://github.com/unslothai/llama.cpp/pull/251/commits/2f5ad7f7908d2bd655a82faf2f7b5b9a984f787f", + "https://github.com/unslothai/llama.cpp/pull/251/commits/3d6eeaf629dfdadd47a292d58cb1e6e52bda0120", "https://github.com/unslothai/llama.cpp/pull/270/commits/7968eda613e427ad5ccbab61df43246a9a82c53a", "https://github.com/unslothai/llama.cpp/pull/264/commits/5bb68456bc34238cedb50c8683495c6d6f924177", "https://github.com/unslothai/llama.cpp/pull/266/commits/88035b0f74d0fd076df2012a957950fee0c4615f" From 824bfca564986203d28e9129aa1ff4cf3c14b2c9 Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Wed, 14 Oct 2026 12:59:09 -0700 Subject: [PATCH 06/11] pr-set: repin unslothai#240 (target backends for a draft that borrows target tensors) --- scripts/unsloth/pr-set.json | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/scripts/unsloth/pr-set.json b/scripts/unsloth/pr-set.json index 92f573171f57..5e12fd77318e 100644 --- a/scripts/unsloth/pr-set.json +++ b/scripts/unsloth/pr-set.json @@ -28,7 +28,7 @@ "https://github.com/unslothai/llama.cpp/pull/149/commits/efa8b0e689ac8ba13f5840d9c26aa227ffd07153", "https://github.com/unslothai/llama.cpp/pull/176/commits/e2519ebc09468b477847a7529834a8da998ed2da", "https://github.com/unslothai/llama.cpp/pull/239/commits/bec216444552a20bc39bf2156fa5396154970608", - "https://github.com/unslothai/llama.cpp/pull/240/commits/d9dd1810da0075ff0000a352055ab246b2738d5e", + "https://github.com/unslothai/llama.cpp/pull/240/commits/235c558d2377ec373d45071d6943851cd5e26de9", "https://github.com/unslothai/llama.cpp/pull/241/commits/20dd977eae5e1a3b0633bba85f03d8b09692379f", "https://github.com/unslothai/llama.cpp/pull/243/commits/f427abb1c8b0cc66c138fa1af42c9eaa368eec82", "https://github.com/unslothai/llama.cpp/pull/251/commits/3d6eeaf629dfdadd47a292d58cb1e6e52bda0120", From 57d259b18315df834093d93af3aaf8c13be8c01f Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Wed, 14 Oct 2026 14:03:04 -0700 Subject: [PATCH 07/11] pr-set: repin unslothai#241 (CUDA graph key for a graph with no nodes) --- scripts/unsloth/pr-set.json | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/scripts/unsloth/pr-set.json b/scripts/unsloth/pr-set.json index 5e12fd77318e..8b1aae7ad28b 100644 --- a/scripts/unsloth/pr-set.json +++ b/scripts/unsloth/pr-set.json @@ -29,7 +29,7 @@ "https://github.com/unslothai/llama.cpp/pull/176/commits/e2519ebc09468b477847a7529834a8da998ed2da", "https://github.com/unslothai/llama.cpp/pull/239/commits/bec216444552a20bc39bf2156fa5396154970608", "https://github.com/unslothai/llama.cpp/pull/240/commits/235c558d2377ec373d45071d6943851cd5e26de9", - "https://github.com/unslothai/llama.cpp/pull/241/commits/20dd977eae5e1a3b0633bba85f03d8b09692379f", + "https://github.com/unslothai/llama.cpp/pull/241/commits/7fa42ff5efb5388ed8f47cec7bce4f8fa7dfe546", "https://github.com/unslothai/llama.cpp/pull/243/commits/f427abb1c8b0cc66c138fa1af42c9eaa368eec82", "https://github.com/unslothai/llama.cpp/pull/251/commits/3d6eeaf629dfdadd47a292d58cb1e6e52bda0120", "https://github.com/unslothai/llama.cpp/pull/270/commits/7968eda613e427ad5ccbab61df43246a9a82c53a", From e748c4df9be141d7b27e06f639813ac2a58e3c7b Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Wed, 14 Oct 2026 14:07:51 -0700 Subject: [PATCH 08/11] pr-set: repin unslothai#251 (measure a shared context at the placement being tried) --- scripts/unsloth/pr-set.json | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/scripts/unsloth/pr-set.json b/scripts/unsloth/pr-set.json index 8b1aae7ad28b..c4bf7d665a2e 100644 --- a/scripts/unsloth/pr-set.json +++ b/scripts/unsloth/pr-set.json @@ -31,7 +31,7 @@ "https://github.com/unslothai/llama.cpp/pull/240/commits/235c558d2377ec373d45071d6943851cd5e26de9", "https://github.com/unslothai/llama.cpp/pull/241/commits/7fa42ff5efb5388ed8f47cec7bce4f8fa7dfe546", "https://github.com/unslothai/llama.cpp/pull/243/commits/f427abb1c8b0cc66c138fa1af42c9eaa368eec82", - "https://github.com/unslothai/llama.cpp/pull/251/commits/3d6eeaf629dfdadd47a292d58cb1e6e52bda0120", + "https://github.com/unslothai/llama.cpp/pull/251/commits/d466df9211feb3e37d8163ebc6d17bc1be8af189", "https://github.com/unslothai/llama.cpp/pull/270/commits/7968eda613e427ad5ccbab61df43246a9a82c53a", "https://github.com/unslothai/llama.cpp/pull/264/commits/5bb68456bc34238cedb50c8683495c6d6f924177", "https://github.com/unslothai/llama.cpp/pull/266/commits/88035b0f74d0fd076df2012a957950fee0c4615f" From 7ea5eaf977c1982251bb2913e1e2515b11226ef0 Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Wed, 14 Oct 2026 14:12:17 -0700 Subject: [PATCH 09/11] pr-set: repin unslothai#240 (vocab-only models skip the shared MTP head check) --- scripts/unsloth/pr-set.json | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/scripts/unsloth/pr-set.json b/scripts/unsloth/pr-set.json index c4bf7d665a2e..bd408d08d3e1 100644 --- a/scripts/unsloth/pr-set.json +++ b/scripts/unsloth/pr-set.json @@ -28,7 +28,7 @@ "https://github.com/unslothai/llama.cpp/pull/149/commits/efa8b0e689ac8ba13f5840d9c26aa227ffd07153", "https://github.com/unslothai/llama.cpp/pull/176/commits/e2519ebc09468b477847a7529834a8da998ed2da", "https://github.com/unslothai/llama.cpp/pull/239/commits/bec216444552a20bc39bf2156fa5396154970608", - "https://github.com/unslothai/llama.cpp/pull/240/commits/235c558d2377ec373d45071d6943851cd5e26de9", + "https://github.com/unslothai/llama.cpp/pull/240/commits/d2dd44f2c60e6c7d4c9df5bac8a95f690cd2c776", "https://github.com/unslothai/llama.cpp/pull/241/commits/7fa42ff5efb5388ed8f47cec7bce4f8fa7dfe546", "https://github.com/unslothai/llama.cpp/pull/243/commits/f427abb1c8b0cc66c138fa1af42c9eaa368eec82", "https://github.com/unslothai/llama.cpp/pull/251/commits/d466df9211feb3e37d8163ebc6d17bc1be8af189", From 1fa1fd65cabd187ee880fc5df3d8d3d9f67dd821 Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Wed, 14 Oct 2026 14:27:30 -0700 Subject: [PATCH 10/11] pr-set: repin unslothai#95 (penalties skip history ids outside the candidates) --- scripts/unsloth/pr-set.json | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/scripts/unsloth/pr-set.json b/scripts/unsloth/pr-set.json index bd408d08d3e1..dbfc35d36ae8 100644 --- a/scripts/unsloth/pr-set.json +++ b/scripts/unsloth/pr-set.json @@ -22,7 +22,7 @@ "https://github.com/ggml-org/llama.cpp/pull/25731/commits/607b193a546bf7ebd17cca1c3e355c622fcef605", "https://github.com/unslothai/llama.cpp/pull/70/commits/883f2c9ba78f3847148454adf025da29385fff3e", "https://github.com/unslothai/llama.cpp/pull/61/commits/02a015c22ad0f9a6a2b1d4361900b45aa212f936", - "https://github.com/unslothai/llama.cpp/pull/95/commits/3db8cb5b2e9bf291057b9f19960e8601a162da81", + "https://github.com/unslothai/llama.cpp/pull/95/commits/496a8a646f19b244b5e2eaf4cf64f920ed31c1a3", "https://github.com/unslothai/llama.cpp/pull/158/commits/fb9b9462041ad58cde9a274e7984c3c6fb01b279", "https://github.com/unslothai/llama.cpp/pull/157/commits/6c6da89266ba7839d825c9997782af4f4d26b81b", "https://github.com/unslothai/llama.cpp/pull/149/commits/efa8b0e689ac8ba13f5840d9c26aa227ffd07153", From f80003caf39f7456cc0f0edc87638f02599c77fd Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Wed, 14 Oct 2026 14:44:02 -0700 Subject: [PATCH 11/11] pr-set: repin unslothai#61 (IQ1_XS types in the GCN and older Pascal MMQ tables, BF16 converters) --- scripts/unsloth/pr-set.json | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/scripts/unsloth/pr-set.json b/scripts/unsloth/pr-set.json index dbfc35d36ae8..3f41dfe7b629 100644 --- a/scripts/unsloth/pr-set.json +++ b/scripts/unsloth/pr-set.json @@ -21,7 +21,7 @@ "https://github.com/ggml-org/llama.cpp/pull/24423/commits/6a7388aac74ab896d9d4f0fb187ddac4847c02f0", "https://github.com/ggml-org/llama.cpp/pull/25731/commits/607b193a546bf7ebd17cca1c3e355c622fcef605", "https://github.com/unslothai/llama.cpp/pull/70/commits/883f2c9ba78f3847148454adf025da29385fff3e", - "https://github.com/unslothai/llama.cpp/pull/61/commits/02a015c22ad0f9a6a2b1d4361900b45aa212f936", + "https://github.com/unslothai/llama.cpp/pull/61/commits/05a66933fc06f6b2eda2288caf506e09f400ed2b", "https://github.com/unslothai/llama.cpp/pull/95/commits/496a8a646f19b244b5e2eaf4cf64f920ed31c1a3", "https://github.com/unslothai/llama.cpp/pull/158/commits/fb9b9462041ad58cde9a274e7984c3c6fb01b279", "https://github.com/unslothai/llama.cpp/pull/157/commits/6c6da89266ba7839d825c9997782af4f4d26b81b",