the short version
- ggml v1.0.0 is a breaking release. It prunes the deprecated public API and removes the DIAG_MASK_INF and DIAG_MASK_ZERO ops, so code that calls them has to change.
- The prune bumps the RPC protocol so that incompatible peers are rejected, and a separate change moves RPC to protocol v9. Every node in a ggml RPC setup has to be upgraded together.
- The release notes make no explicit API or ABI stability promise and do not mention GGUF. Binding authors should not treat 1.0 as a frozen interface yet.
- Several numerical behaviors changed: F32 src1 for CPU MUL_MAT, F32 accumulation of FlashAttention V on CPU, and deterministic tie-breaking in CUDA/Metal radix Top-K. Output-matching tests may need new baselines.
ggml v1.0.0, released on GitHub on 13 November, is a breaking release. It prunes the deprecated public API, removes the legacy DIAG_MASK_INF and DIAG_MASK_ZERO ops, and bumps the RPC protocol, now at v9, so that incompatible peers are rejected. The release notes make no stated commitment on API stability, ABI compatibility or GGUF. Most of the remaining changelog since v0.29.0 adds FP8 support, configurable YaRN truncation, and performance and correctness work across CUDA, Metal, Vulkan, SYCL, OpenCL, WebGPU, Hexagon, CPU and smaller backends.
If you only consume ggml through a pinned llama.cpp build, most of this arrives without any action on your part. The changes carry llama/ PR numbers and were synced into ggml from llama.cpp. The exposure is greater for anyone who calls ggml directly: language bindings, forks, whisper.cpp-style applications, and anyone running ggml RPC across several machines. For them, the 1.0.0 tag is mainly a set of removed symbols and a protocol bump to plan around.
What ggml v1.0.0 removes from the API
The release summary says ggml 1.0.0 'prunes the deprecated public API and the legacy DIAG_MASK ops'. The specific changes are llama/29801 (the prune) and llama/31946. The changelog also includes a commit titled 'examples : adapt to removal of ggml_diag_mask_inf', so the ggml repository's own examples needed edits to keep building. Downstream code that builds causal masks with ggml_diag_mask_inf will need the same treatment. The release notes do not list every deprecated function that was pruned, so binding authors should diff the public headers between v0.29.0 and v1.0.0 before upgrading.
The release also adds API. ggml_rope_set_truncate and ggml_rope_get_truncate control rounding of the YaRN correction bounds (llama/30206, llama/30802). The backward pass for TRI is now implemented (llama/30712). MUL_MAT src1 precision now has tests and gating for ggml_prec_set_acc (llama/31896, llama/31851). The notes list YaRN truncation config support for OpenVINO, SYCL, Vulkan, WebGPU and RPC under llama/30802. They do not state backend coverage for this setting beyond that list.
RPC protocol v9 forces lockstep upgrades
Two changes affect RPC. The prune bumps the RPC protocol 'to reject incompatible peers'. A separate change moves the protocol to v9 with faster loading and graceful failure on unreachable endpoints or failed handshakes (llama/29413, llama/30415). In practice, a v1.0.0 node and an older peer are not expected to interoperate, and the documented behavior is that a failed handshake fails gracefully. Teams running ggml RPC across several machines should schedule a coordinated rollout and keep the older build available for rollback on every node, not only on the client.
Numerical changes that may break golden-output tests
Several changes alter results without any change to the API:
The CPU backend now uses F32 src1 for MUL_MAT (llama/31854) and accumulates FlashAttention V in F32 (llama/29810). CUDA and Metal radix Top-K now break ties deterministically (llama/31084). Metal TRI now uses a select instead of scaling by the comparison (llama/30711). CUDA MoE selection sanitizes post-bias NaNs (llama/31005). SYCL now rounds exactly when copying to Q4_1, Q5_1 and Q8_0 (llama/31465).
These changes target determinism, precision or rounding correctness. Each can also shift logits enough to fail a byte-for-byte regression test recorded on v0.29.0. Plan to regenerate baselines and to compare perplexity or task accuracy, not exact outputs. The release notes do not quantify the size of any of these numerical shifts.
Backend support added since v0.29.0
- CUDA: few-row MMQ for batch sizes 2-24, MMVF for BF16 few-row matmuls, OCP FP8 support, fused NVFP4 gate/up matmuls with GLU, an sm70 head-size-256 FlashAttention config, MMA head sizes 40/72, and a d_state=64 SSM_SCAN fast path. HIP gets a fixed BLAS workspace and an MFMA lightning indexer for CDNA2.
- Metal: fused MoE routing, fusable node groups kept inside one command buffer, mul_mm for 32<=K<64, and out-of-bounds fixes in the FA output and SOLVE_TRI kernels.
- Vulkan: an Intel FlashAttention prefill kernel, ReBAR detection, batched tensor copies, RDNA3/4 large-tile split-K, MoltenVK out-of-bounds fixes, and scalar FlashAttention capped at 8 rows for large heads on Qualcomm.
- SYCL: a device-side MoE scheduler, a thread-safe oneDNN FA cache, rejection of buffers that would exceed VRAM, and an AOT fix for allocations above 4GB.
- OpenCL: decomposed Adreno FlashAttention, q6_k tiled, q8_0 and f32/q5_k multi-column GEMV kernels, and multi-column GEMV enabled by default for Adreno X2.
- WebGPU, Hexagon, CANN, ZDNN, OpenVINO: WebGPU adds GROUP_NORM, MEAN, COUNT_EQUAL and PAD_REFLECT_1D. Hexagon adds Q5_0/Q5_1, conv_3d and automatic libhtp loading. CANN fails early on unsupported buffer types. ZDNN accepts non-zdnn tensors. OpenVINO supports the YaRN truncation config.
In the core, the scheduler no longer uploads inputs that are not read, and it prefetches only inputs produced by earlier splits (llama/31419, llama/32100). One commenter on Hacker News said that when an MoE model slightly exceeds GPU memory, llama.cpp and ggml can now cache less-used experts in host memory instead of failing to load. The release notes describe a narrower change: the MoE cache 'can compute a share of the cache misses on the CPU and copy the results back asynchronously' (llama/30966), in a commit titled 'llama_moe_cache: run missed experts on CPU, add tuning script'. Treat that as the documented behavior, not a general guarantee that oversized models will load.
One commenter reported a 5 to 10 percent speedup with each roughly monthly llama.cpp upgrade. They cited a 50K-token run that dropped from 4:22 to 3:59, and a second commenter said that matched their own CPU-only results. These are user anecdotes about llama.cpp upgrades in general, not a benchmark of this release, and the release notes give no performance figures.
Version 1.0 does not state an ABI or GGUF guarantee
One commenter on Hacker News asked whether 1.0 comes with a stability commitment for the public C API, because that would make in-process Python bindings for llama.cpp easier to build. The release notes do not answer the question. They contain no statement about API stability until 2.0, nothing about ABI compatibility for shared-library consumers, and no mention of GGUF or of whether existing model files are affected. The notes link to a page titled 'Releases and versioning of ggml-org projects', but its contents are not part of the release text.
Until that policy is read and confirmed, binding authors should keep doing what they do now. Pin an exact ggml commit, regenerate FFI declarations from headers on every bump, and do not assume that a 1.x minor release will be drop-in compatible.
Upgrade steps for llama.cpp, whisper.cpp and bindings
Most commits in the changelog carry llama/ PR numbers and arrive through 'sync : llama.cpp' commits, which indicates the changes landed in llama.cpp first. Some llama-side adaptations appear directly, such as 'llama, mtmd: follow the reference precision of ggml_prec_set_acc calls' (llama/31851). The changelog also shows a multimodal refactor in which clip.h is removed and its code consolidated into mtmd (llama/31865). The sources say nothing about whisper.cpp specifically, so the steps below cover only what the ggml changes imply for any direct consumer.
- Diff the ggml public headers between v0.29.0 and v1.0.0, and remove or replace every call to a pruned symbol, starting with ggml_diag_mask_inf.
- Replace legacy DIAG_MASK_INF and DIAG_MASK_ZERO graph construction with whatever masking path your upstream (llama.cpp or the ggml examples) now uses.
- If you use RPC, upgrade every rpc-server and client together, and confirm that handshake failures surface as clean errors in your orchestration.
- If you include clip.h from llama.cpp's multimodal code, move to the consolidated mtmd interface.
- Regenerate numerical baselines on CPU, CUDA, Metal and SYCL because of the F32 src1, F32 FA V, deterministic Top-K and exact-rounding changes.
- If your models use YaRN, decide whether to set truncation explicitly with ggml_rope_set_truncate, and check that it behaves the same on each backend you ship.
Questions this raises
What does ggml v1.0.0 remove?
It prunes the deprecated public API and removes the legacy DIAG_MASK_INF and DIAG_MASK_ZERO ops. The release notes do not list every pruned function, so binding authors should diff the public headers between v0.29.0 and v1.0.0.
Can a ggml v1.0.0 RPC node talk to older peers?
It is not expected to. The protocol is now v9 and is bumped to reject incompatible peers, and a failed handshake fails gracefully. Multi-machine setups should upgrade every node together and keep the older build on hand for rollback.
Does ggml 1.0 promise API, ABI or GGUF stability?
The release notes make no such commitment and do not mention GGUF. They link to a 'Releases and versioning of ggml-org projects' page, but until that policy is confirmed, pin an exact commit and regenerate FFI declarations on every bump.
Spotted an error? Email samir.s.sengupta@gmail.com and I will correct it and date the change at the top of the note. How these notes are researched, drafted and checked is in the editorial policy.
