hotfix: lean ROCm archive so b10018-1.2.0 can publish - #53
Merged
Conversation
The ROCm archive hit 9.6 GB and blew past GitHub's 2 GB asset limit, so publish-release 422'd and b10018-1.2.0 failed to publish. Cause: I bundled the whole rocBLAS/hipBLASLt Tensile kernel database (kernels for every gfx arch) to make the archive self-contained. Drop it: ship binaries + libggml-hip.so (~0.45 GB) only, require the system ROCm runtime like upstream's ROCm builds. Archive drops ~9.6 GB -> ~0.6 GB. Adds a README-ROCm.txt noting the runtime requirement.
NighmareGit
added a commit
to NighmareGit/atomic-llama-cpp-turboquant
that referenced
this pull request
Aug 1, 2026
Port the concept of ikawrakow@f43a9f1 (per-byte CUDA MoE offload threshold) as a clean env-var-gated change. For GGML_OP_MUL_MAT_ID (MoE expert matmul), scale the offload threshold by weight row size so smaller weight formats offload at proportionally smaller batch sizes — bandwidth cost is lower, so offloading is worthwhile earlier. Env var: GGML_OP_OFFLOAD_MIN_BATCH_PER_BYTE (default -1 = disabled). Zero-risk to the default path. Files: ggml/src/ggml-cuda/ggml-cuda.cu (+21/-1) Reference: .scratch/research/steal-rework-feasibility.md (LEDGER AtomicBot-ai#53)
NighmareGit
added a commit
to NighmareGit/atomic-llama-cpp-turboquant
that referenced
this pull request
Aug 1, 2026
Add a runtime verification hook in graph_mtp that logs whether the TurboQuant KV rotation matrices are visible to the MTP draft head. Finding: the MTP path calls the shared build_attn(), which applies TQ WHT rotation gated purely on k->type == GGML_TYPE_TURBO3_0/4_0/2_0. Because MTP shares the target model's KV cache context, TQ rotation engages automatically — no MTP-specific hook is required. The log line confirms at runtime that the MTP path sees the same turbo_rot / innerq tensors as the target, proving the path is wired correctly. A WARN is emitted if the target has TQ rotation but the MTP k_rot is NULL (would indicate a cache mismatch). Verify status: build-clean (ROCm/HIP, gfx1100). Functional 7900 test deferred — GPU busy with VVRAM eviction agent live-testing. Files: src/models/qwen35moe.cpp (+22/-0) Reference: .scratch/research/steal-rework-feasibility.md (LEDGER AtomicBot-ai#53)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
b10018-1.2.0 built all 9 backends successfully but publish-release 422'd: the ROCm .zip exceeded GitHub's 2 GB asset limit (archive was 9.6 GB).
Root cause: the ROCm archive bundled the full rocBLAS/hipBLASLt Tensile kernel database (precompiled GEMM kernels for every gfx arch) to be self-contained — that's multi-GB. Stock llama.cpp ROCm builds don't do this.
Fix: lean archive — binaries +
libggml-hip.so(~0.45 GB) only, require the system ROCm runtime (documented in a bundled README-ROCm.txt). Archive drops from ~9.6 GB to ~0.6 GB, well under the limit.Admin-merge this, then I delete + re-push the
b10018-1.2.0tag to rebuild and publish the release (with ROCm). CI-only change; I'll sync dev right after.