ggml-metal: fix int32 overflow in kernel_mul_mm batched dst offsets (deep-context corruption) - #28210

Closed

feni6 wants to merge 63 commits into

ggml-org:masterfrom

feni6:metal-mul-mm-u64-dst-offsets

Commits on Aug 26, 2026

llama: load glm5next hparams and tensors

danielhanchen
authored and
danielhanchen
committed
a505a75
llama: add glm5next mHC wide residual

danielhanchen
authored and
danielhanchen
committed
9e1cc20
llama: add glm5next KDA linear attention
danielhanchen
committed
20a8077
llama: add glm5next MoE feed-forward with clamped SwiGLU
danielhanchen
committed
1d99a5a
llama: add glm5next dense DSA attention
danielhanchen
committed
f320a28
llama: save the indexer kpool and hyper-connection keys
danielhanchen
committed
7f2560e
test-llama-archs: enable glm5next
danielhanchen
committed
cd69d60
llama: assert the glm5next indexer selection width
danielhanchen
committed
6a58359
glm5next: trim DSA dense comments
danielhanchen
committed
3841592
llama: third kv cache for the glm5next lightning indexer
danielhanchen
committed
2db8295
glm5next: trim pooled indexer comments
danielhanchen
committed
839597c
glm5next: lightning indexer graph, pooled sparse selection

danielhanchen
authored and
danielhanchen
committed
8c49836
mtmd: don't normalize patch embeddings when norm_embd is absent
danielhanchen
committed
fe95953
convert: refuse to write a GGUF with zero tensors
danielhanchen
committed
582fe81
mtmd: add glm5next (GLM-5.3-Flash) vision tower
danielhanchen
committed
9901ab4
mtmd: implement the glm5next 0826 image preprocessor
danielhanchen
committed
41cd923
mtmd: resample glm5next images with bicubic, matching the reference
danielhanchen
committed
02fa4a4
llama: enable ignore_merges for the glm4 pre-tokenizer

danielhanchen
authored and
danielhanchen
committed
81e3e67
llama: shorten comment
danielhanchen
committed
2d9570d
glm5next: repair conflict resolution from the rebase onto master
danielhanchen
committed
6c59f9b

Commits on Aug 27, 2026

glm5next: do not inherit deepseek4's n_embd_out
danielhanchen
committed
eab9ee9
glm5next: keep precision-sensitive tensors at source precision
danielhanchen
committed
869e878
glm5next: restore the vision preprocessor tests
danielhanchen
committed
282ef61
glm5next: fix E301 lint in gguf_writer
danielhanchen
committed
cadbe97
glm5next: pool the indexer per sequence so a unified KV cache works
danielhanchen
committed
204fa70
glm5next: test the per-sequence pool runs instead of the old refusal
danielhanchen
committed
ef531f8
glm5next: store the indexer selection masks in f16

danielhanchen
authored and
danielhanchen
committed
e88c92d
glm5next : use the fused lightning indexer for the pool scores

danielhanchen
authored and
danielhanchen
committed
2e0e57f

Commits on Aug 28, 2026

Commits on Aug 30, 2026

Faster inference
danielhanchen
committed
0069971
Add MTP support
danielhanchen
committed
d07e71e

Commits on Sep 1, 2026

tests: add glm5next --depth-sweep mode (CPU-vs-device logit divergence over long context)
Your Name
committed
238f880
tests: depth-sweep gains --topk override and same-backend ubatch A/B (--ub2/--b-cpu)
Your Name
committed
ed1313c
tests: depth-sweep --iheads/--iklen overrides (real glm5next: 32 heads, key 128)
Your Name
committed
2841c7b
tests: depth-sweep --layers/--dlead (glm5next layer count + leading dense blocks)
Your Name
committed
40f7a34
ggml-metal: 64-bit output offsets in mul_mm batched dst indexing
Your Name
committed
f5d5216
llama-context: env-gated alloc-map dump (LLAMA_ALLOCDUMP_*) for wrap-landing investigation
Your Name
committed
c31aa2d
metal: gpuAddress diagnostic (codex patch) + allocdump gpu_data/gpu_base columns
Your Name
committed
b439462
allocdump: LLAMA_ALLOCDUMP_REQ_TOKENS full-ubatch filter, bidirectional stride (grokk-reply)
Your Name
committed
42b9e5e
metal: LLAMA_METAL_CAPTURE_ARM_FILE one-shot capture arming + tensor-named debug groups (speed conference 008/009)
Your Name
committed
b5d5264
graph: env-gated packed-MQA decode transform (speed-conf 019 Patch A); tests: fix sweep-globals order, add --heads override
Your Name
committed
b5fe910
graph: split packed-MQA modes (kq/kqv/both) + f32 prec on packed KQV, historical cb names (grokk 020); tests: DSA n_head_kv=1 under --heads, FA off in harness (production parity)
Your Name
committed
d33b0ab
tests: scope depth-sweep FA-off via arg (other arch tests keep AUTO); graph: NON_FA branch marker (codex 026 dual-marker doctrine)
Your Name
committed
19446fe
graph: strict packed-MQA mode parse - numeric 2 no longer aliases to both (grokk 028), unrecognized values warn and disable
Your Name
committed
5f3d764
kv-cache/graph: Patch B — derived transposed MLA-V mirror (LLAMA_MLA_V_MIRROR); contracts 4.1-4.6+7; hybrid-k wrapper fills v_idxs (the sub-input set_input is bypassed there); trunk-sparse consumer…
Your Name
committed
37a483c
graph/kv-cache: Patch B hardening (grokk 033 2.1/2.3/2.4 + codex 034 2-4): dense build_attn mirror wiring, wrapper can_reuse width check, strict V parser + MLA_V_MIRROR ACTIVE marker, construction …
Your Name
committed
c4cc25e
graph: exact v_idxs width check in attn_k sub-input; name mirror update v_mirror_upd in both consumers (grokk 036)
Your Name
committed
f86a0da
kv-cache: scope V-mirror opt-in to nope-only GLM5Next (codex 051 s3)

Your Name
and
claude
committed
0519ed8

Commits on Sep 2, 2026

tests: single-token decode step in depth-sweep so packed decode branches are exercised (grokk 054 s3)

Your Name
and
claude
committed
dd80b05
tests: --dump-1tok writes the decode-1 device+cpu logits for cross-run oracles

Your Name
and
claude
committed
92177ae
glm5next: D0 gathered DSA decode (speed conference 056/057) — LLAMA_GLM5_GATHERED_DSA

Your Name
and
claude
committed
fbc3542
glm5next/tests: D0 P0 batch (codex 059 / grokk 060)

Your Name
and
claude
committed
4cbe31e
glm5next: PAD32 — pad gathered slots to the mul_mm tile (codex 063 / grokk 064)

Your Name
and
claude
committed
6056d05
graph: Patch C view-only — drop redundant conts on packed-path singleton permutes (codex 043/051, grokk 045)

Your Name
and
claude
committed
c4ec35f
graph/context: R0 — kpool can_reuse contract + split host timers (codex 068/072, grokk 070/073)

Your Name
and
claude
committed
048799f
graph/context: E0a — live certificate-subsumption invariant, host-side class compare

Your Name
and
claude
committed
1f9064b
context/kpool: timer + reuse-contract repairs (grokk 076/078, codex 077/079)

Your Name
and
claude
committed
37447da
context: rebuild event ledger replaces nonzero percentiles (grokk 081 s2, codex 082 s3)

Your Name
and
claude
committed
03dc115
glm5next: E0b+c MASKLESS gathered decode (codex 066, grokk 067) + diamond PASS

Your Name
and
claude
committed
18dcfb3