From 40ac0cb9ac6afe061eb125531d6f67db3243382a Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Fri, 9 Oct 2026 05:11:51 -0700 Subject: [PATCH 1/4] unsloth: pin #241 again, now that it carries the CUDA graph cache cap removal The 64 entry cap came from #241's own shape keyed graph cache commit, so the fix belongs on #241. Its branch now includes #253's commit (same head, 20dd977), and the pin goes back to #241. The built tree is unchanged. --- scripts/unsloth/feature-checks.json | 2 +- scripts/unsloth/pr-set.json | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/scripts/unsloth/feature-checks.json b/scripts/unsloth/feature-checks.json index 6db05da5e4fd..9a404223ed9f 100644 --- a/scripts/unsloth/feature-checks.json +++ b/scripts/unsloth/feature-checks.json @@ -99,7 +99,7 @@ "unslothai#152": "per-run mmap of a context's tensors; a memory-layout change with no observable output difference", "unslothai#157": "cudaMemcpyDefault in the ggml_cuda_cpy 2D fast path; needs a CUDA host", "unslothai#158": "ROCm_Host compute buffer type on HIP integrated GPUs; needs a ROCm host", - "unslothai#253": "carries #241 (CUDA graph cache key, macOS row prefetch and the F16 glm5-next indexer cache under a quantized -ctk (the NextN draft graph is upstream since ggml-org#29928)) and drops the CUDA graph cache count cap so tensor split keeps replaying its 2 * n_layers + 1 graphs per device; the first needs a CUDA host, the second has no observable output difference, the last needs a real GLM-5-Next model; the cap removal also needs a multi-GPU CUDA host", + "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" } diff --git a/scripts/unsloth/pr-set.json b/scripts/unsloth/pr-set.json index 2d8e6cc86588..d793b6033a27 100644 --- a/scripts/unsloth/pr-set.json +++ b/scripts/unsloth/pr-set.json @@ -30,7 +30,7 @@ "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/36175a8325e24303704cfbab1b7d8a0bd30500a5", - "https://github.com/unslothai/llama.cpp/pull/253/commits/20dd977eae5e1a3b0633bba85f03d8b09692379f", + "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/13f4c714dff3bf7aa1a83be7490102b7706244ee" ] From ba64bcf97f759de6da8d81ee6d8f72eaedd28efb Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Fri, 9 Oct 2026 06:06:34 -0700 Subject: [PATCH 2/4] unsloth: repin ggml-org#25731 onto b11515 b11515 brings ggml-org#30147, which loops the CUDA PAD kernel over rows and slices past 65535. Inkling's int64 dst_idx fix touched the same line, so the pin stopped merging. The Inkling branch now merges b11515 and keeps the cast inside the new loop (test-backend-ops -o PAD: 37/37 on CUDA). --- 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 d793b6033a27..ffaffeba3159 100644 --- a/scripts/unsloth/pr-set.json +++ b/scripts/unsloth/pr-set.json @@ -19,7 +19,7 @@ ], "prs": [ "https://github.com/ggml-org/llama.cpp/pull/24423/commits/643189144c98fba431e8d2da169210b7b6d5d930", - "https://github.com/ggml-org/llama.cpp/pull/25731/commits/fddee40253015ff5f7e578fa941bd074d591cf33", + "https://github.com/ggml-org/llama.cpp/pull/25731/commits/343a03721776836a4b200be18c7fe220dd5c9f20", "https://github.com/unslothai/llama.cpp/pull/70/commits/883f2c9ba78f3847148454adf025da29385fff3e", "https://github.com/unslothai/llama.cpp/pull/61/commits/ef45f212c90d5769e0dd8d482f99fd66760312e7", "https://github.com/unslothai/llama.cpp/pull/95/commits/3db8cb5b2e9bf291057b9f19960e8601a162da81", From 8b2eaaa962a68abd3cf42a5f957d183a5b4d9eb1 Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Fri, 9 Oct 2026 08:35:41 -0700 Subject: [PATCH 3/4] unsloth: repin #61 onto b11517 b11517 brings ggml-org#30168, which makes the src1 precision an explicit argument of the CUDA MMQ helpers and drops its default. #61's three IQ1 types now pass GGML_PREC_Q8 in the switch cases, the tile loaders and mul_mat_q_case, as upstream does for IQ1_S (test-backend-ops MUL_MAT, MUL_MAT_ID, GET_ROWS on CUDA all pass, incl. the MMQ batch sizes). --- 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 ffaffeba3159..56437920ddbb 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/643189144c98fba431e8d2da169210b7b6d5d930", "https://github.com/ggml-org/llama.cpp/pull/25731/commits/343a03721776836a4b200be18c7fe220dd5c9f20", "https://github.com/unslothai/llama.cpp/pull/70/commits/883f2c9ba78f3847148454adf025da29385fff3e", - "https://github.com/unslothai/llama.cpp/pull/61/commits/ef45f212c90d5769e0dd8d482f99fd66760312e7", + "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/158/commits/abfc45b9cb21eae4848cb82196e659f42c9a8341", "https://github.com/unslothai/llama.cpp/pull/157/commits/6c6da89266ba7839d825c9997782af4f4d26b81b", From f090ebeed34bd2e0a5d5ad6e21e3484bcd85f450 Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Fri, 9 Oct 2026 09:24:52 -0700 Subject: [PATCH 4/4] unsloth: repin ggml-org#25731 onto b11529 b11529 brings ggml-org#28636, which moves the ModelOpt FP8 dequant loop in conversion/base.py into a dequant_fp8() helper. Inkling's stacked-expert NVFP4 rejection sat in front of that loop; it now sits in front of the call. --- 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 56437920ddbb..c0606c24edca 100644 --- a/scripts/unsloth/pr-set.json +++ b/scripts/unsloth/pr-set.json @@ -19,7 +19,7 @@ ], "prs": [ "https://github.com/ggml-org/llama.cpp/pull/24423/commits/643189144c98fba431e8d2da169210b7b6d5d930", - "https://github.com/ggml-org/llama.cpp/pull/25731/commits/343a03721776836a4b200be18c7fe220dd5c9f20", + "https://github.com/ggml-org/llama.cpp/pull/25731/commits/31940fbb55400a2e4b28ffaaf8cf20c358e14094", "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",