Repository navigation

hip/gfx1151: fix mmap load hang, enable QSA get_rows rows %32, long-context session support - #166

Closed

gaetan-puleo wants to merge 398 commits into

rocm/longctx-stability

Closed

hip/gfx1151: fix mmap load hang, enable QSA get_rows rows %32, long-context session support#166
gaetan-puleo wants to merge 398 commits into
masterfrom
rocm/longctx-stability

This pull request is big! We're only showing the most recent 250 commits

Commits on Sep 8, 2026
Commits on Sep 11, 2026
- Merge pull request #42 from ajaxdude/jeromecoste-microsoft-small-project-contribution
  Laurent Zuijdwijk
  authored
  6b1f308
- Merge remote-tracking branch 'pr/27' into resolve-pr-27
  Laurent Zuijdwijk
  committed
  29e091e
Commits on Sep 12, 2026
- merge: sync Halo master while preserving Strix kernels
  Laurent Zuijdwijk
  committed
  054a736
- Merge pull request #35 from arc-uri-el/contribution/checkpoint-io-20260908
  Laurent Zuijdwijk
  authored
  0007bc6
- Merge branch 'master' into sync/halo-master-2026-09-11
  Laurent Zuijdwijk
  authored
  a142a61
- Merge pull request #47 from halo-box/sync/halo-master-2026-09-11
  Laurent Zuijdwijk
  authored
  6548035
Commits on Sep 13, 2026
- Merge pull request #32 from halo-box/rocmfpx/vulkan-2026-09-07
  Laurent Zuijdwijk
  authored
  6994643
Commits on Sep 14, 2026
- server : rename --cache-disk to --cache-dir
  
  Local Dev
  authored and
  local-merge
  committed
  f2431a2
- llama: flatten the GDN output projection input to 2-D
  SimonTeixidor
  committed
  b009b79
- qwen35, qwen35moe: run gated_delta_net in place on the recurrent state cache
  SimonTeixidor
  committed
  e3f31cf
Commits on Sep 15, 2026
- chat : enforce string enums in Qwen XML tool arguments (#26)
  arc-uri-el
  authored
  e20d6f5
- HIP : keep FA head size 192 off the WMMA kernel (#55)
  SimonTeixidor
  authored
  cfe6bb1
- sync upstream llama.cpp commits 2026 09 15 (#28)
  
  authored
  e3120a1
- --ngram-on-disk and lazymode merge (#29)
  dzannotti
  authored
  77024dd
- ggml: recover two upstream commits dropped during the #28 sync (#30)
  
  authored
  8a400d0
- Record upstream sync (content applied via #28, gaps recovered in #30)
  dzannotti
  committed
  f6a7a6c
- docs: add upstream sync procedure (#31)
  dzannotti
  authored
  4bc01d0
- Merge halo-box/llama.cpp master into strix-llama.cpp
  dzannotti
  committed
  dced43c
Commits on Sep 16, 2026
- pwilkin/strix-halo official merge (#63)
  
  pwilkin
  and
  claude
  authored
  0636c9a
Commits on Sep 17, 2026
- Merge strix master (#63) into the halo-box sync
  
  dzannotti
  and
  claude
  committed
  82aed60
- Merge halo-box/llama.cpp master (#31) into strix-llama.cpp
  
  dzannotti
  and
  claude
  committed
  dffbb78
- HIP : route Q6_K through MMQ up to ne11 = 1024 on RDNA 3.5 (#38)
  
  Nyovelt
  and
  claude
  authored
  f86f7a1
- Merge halo-box/llama.cpp master into strix (2026-09-15) (#64)
  dzannotti
  authored
  2ee5fe2
- docs: record why upstream Vulkan sparse FA was deferred (#65)
  
  dzannotti
  and
  claude
  authored
  8c1c282
- fix: guard the qwen4exp PLE prefetch on the model architecture
  
  dzannotti
  and
  claude
  committed
  57f5bde
- cuda: enable the BF16 WMMA matmul path only for qwen4exp
  
  dzannotti
  and
  claude
  committed
  42f7e89
Commits on Sep 20, 2026
- ggml-cuda : round growing pool requests up to a power of two
  
  michal-zurkowski
  and
  claude
  committed
  2167f5f
Commits on Sep 22, 2026
- Merge remote-tracking branch 'yitizi/pr2-standalone' into persistent-prompt-cache
  storm1er
  committed
  89fb1c6
Commits on Sep 24, 2026
- hip : isolate MMB scratch by backend context
  Mushoz
  committed
  1a005ba
- qsa: maskless multi-sequence block selection
  gaetan-puleo
  committed
  27f2476
- llama : fix qsa_n_kv_window crash when n_stream > 1
  gaetan-puleo
  committed
  adb3c50
- qsa: unify scalar visibility gate for single and multi sequence
  gaetan-puleo
  committed
  c3e9ce9
- Merge pull request #83 from Mushoz/fix/rocm-mmb-context-scratch
  gaetan-puleo
  authored
  e7ff2be
- qsa: fix compact graph reuse, shared-block visibility, and 2-D scan cost
  gaetan-puleo
  committed
  d367e26
- hip: port gufo's Flash-Next prefill/decode ideas to qwen4exp on gfx1151
  
  pwilkin
  and
  claude
  committed
  dd601a1
Commits on Sep 25, 2026
Commits on Sep 26, 2026
- qsa : take 1..512 queries in the per-query decode kernel
  
  authored and
  gaetan-puleo
  committed
  9bc06f5
- qwen4exp : use complete-block selection for every ubatch size
  
  authored and
  gaetan-puleo
  committed
  b4be8b7
- qwen4exp : incremental pooled-key update for ubatches up to 127 tokens
  
  authored and
  gaetan-puleo
  committed
  10c2b9e
- Merge pull request #82 from michal-zurkowski/qsa/decode-small-ubatches
  gaetan-puleo
  authored
  0fc7226
- tests : add --mmb on|off to test-backend-ops
  pwilkin
  committed
  4208491
- Merge halo-box master into perf/mmb-qwen4exp-only
  pwilkin
  committed
  c3ee03a
- cuda : zero quantized padding without a synchronous memset
  pwilkin
  committed
  cb7dfec
- Merge pull request #75 from halo-box/perf/mmb-qwen4exp-only
  pwilkin
  authored
  9484e46
- Merge pull request #74 from halo-box/fix/ple-prefetch-arch-guard
  gaetan-puleo
  authored
  52d7e10
Commits on Sep 27, 2026
- hip: select top-k score ties by increasing column index
  gaetan-puleo
  committed
  ee3fbdc
- hip: keep topk-moe fusion when outputs alias router logits
  gaetan-puleo
  committed
  2b6b434
- qwen4exp: own PLE prefetch inputs in the detached thread
  gaetan-puleo
  committed
  9959849
- server : preserve context checkpoints across slot save/restore
  
  Tough-Respawn
  authored and
  ChrisC381
  committed
  f7ff38a
- spec : MTP draft vocabulary subset (--spec-draft-mtp-vocab N)
  
  SimonTeixidor
  authored and
  Laurent Zuijdwijk
  committed
  4024eb1
- spec : make the MTP draft vocabulary subset per context
  
  SimonTeixidor
  authored and
  Laurent Zuijdwijk
  committed
  511bc9c
- tests : per-context MTP draft vocabulary regression test
  
  SimonTeixidor
  authored and
  Laurent Zuijdwijk
  committed
  9fd255c
- ggml-cuda : shape-aware graph cache key
  
  SimonTeixidor
  authored and
  Laurent Zuijdwijk
  committed
  73aa8cf
- tests : check the tested backend's picks in test_top_k_stable_ties
  Laurent Zuijdwijk
  committed
  b6b813d
- Merge pull request #101 from halo-box/split/pr58-mtp-draft-vocab
  Laurent Zuijdwijk
  authored
  72ea598
- Merge pull request #93 from halo-box/fix/qwen4exp-ple-prefetch-lifetime
  gaetan-puleo
  authored
  db3e211
- Merge pull request #95 from halo-box/fix/hip-topk-moe-alias
  gaetan-puleo
  authored
  4bf9d79
- Merge pull request #94 from halo-box/fix/hip-topk-stable-ties
  gaetan-puleo
  authored
  ab6dfc7
- hip: stabilize 128–512 token QSA incremental key updates
  gaetan-puleo
  authored
  063d804
- hip: keep HC residual BF16 marks stable across scheduler split cuts
  gaetan-puleo
  committed
  27c308e
- Merge pull request #96 from halo-box/fix/qwen4exp-qsa-fast-512-on-moe
  gaetan-puleo
  authored
  ec01a7d
Commits on Sep 28, 2026
- cuda: ROCmFPx load/dequant/dispatch paths for HIP builds
  baraxnaxgaming-commits
  committed
  2fd4c2a
- cuda: MMQ tile path for Q4_0_ROCMFP4_FAST (dp4a + WMMA/MFMA layouts)
  baraxnaxgaming-commits
  committed
  b0b7222
- cuda: drop unreferenced rocmfp4 HIP helper files
  baraxnaxgaming-commits
  committed
  3db7cc3
- cuda: gate Q4_0_ROCMFP4_FAST MMQ to validated architectures
  baraxnaxgaming-commits
  committed
  d3a8ab5
- Merge pull request #100 from halo-box/split/pr58-cuda-graph-key
  Laurent Zuijdwijk
  authored
  7a9196d
Commits on Sep 29, 2026
- Merge upstream llama.cpp (2026-09-29)
  dzannotti
  committed
  57fa06b
- convert : drop unused imports in qwen4exp
  dzannotti
  committed
  a1c9358
- Merge pull request #33 from halo-box/sync/upstream-2026-09-29
  dzannotti
  authored
  e571606
- Merge halo-box/llama.cpp master into strix (2026-09-29)
  
  dzannotti
  and
  claude
  committed
  a4a4bfa
- ci : nightly ROCm release builds in the lemonade-sdk/llamacpp-rocm layout
  
  dzannotti
  and
  claude
  committed
  acbfbc8
- ci : release every merge to master, Linux zips by default
  
  dzannotti
  and
  claude
  committed
  7e4f916
- ci : add a Linux Vulkan zip to each release
  
  dzannotti
  and
  claude
  committed
  ea79e75
- Merge pull request #112 from halo-box/ci/rocm-nightly-releases
  gaetan-puleo
  authored
  3deeb33
- hip: optimize Qwen3.8-Flash-Next prefill on Strix Halo (#91)
  gaetan-puleo
  authored
  8388e3f
- Merge master (#91, #112) into the 2026-09-29 halo sync
  
  dzannotti
  and
  claude
  committed
  6f33279
- Merge pull request #98 from ChrisC381/pr-slot-checkpoints
  dzannotti
  authored
  8cd4a97
- scripts : log instead of print in gather_required_libs.py
  
  dzannotti
  and
  claude
  committed
  856df97
- Merge branch 'master' into feat/hip-fp4-mmq
  dzannotti
  authored
  ef338d7
- ggml : backport persistent view initialization after buffer-size splits (#37)
  
  arc-uri-el
  and
  Yoshi4470
  authored
  12643af
Commits on Sep 30, 2026
Commits on Oct 1, 2026
Commits on Oct 2, 2026
- hip : drop the F32 HC norm output only when the fused gate mix takes its stream mix
  iTechMedic
  committed
  5295c00
- tests : MMB_QUANT_HC when the fused gate mix refuses: shapes, gate weight types, reader rule
  iTechMedic
  committed
  5b2effe
- hip : default build target to gfx1151
  gaetan-puleo
  committed
  4a8c819
- Merge pull request #134 from gaetan-puleo/build/default-hip-target-gfx1151
  gaetan-puleo
  authored
  0d08f3f
- server : /dashboard live stats page and /dashboard/stats endpoint
  
  Laurent Zuijdwijk
  and
  claude
  committed
  a3646f6
- server : dashboard reports lazy_mode (strix) instead of ple_on_disk
  
  Laurent Zuijdwijk
  and
  claude
  committed
  51dc1da
- server : link the chat UI and the /dashboard page to each other
  
  Laurent Zuijdwijk
  and
  claude
  committed
  e8b4cfd
- Merge pull request #131 from iTechMedic/feat/qwen4exp-small-pp-master
  gaetan-puleo
  authored
  18a46bd
- Merge master 18a46bda4 into fix/hc-gate-mix-gemm-eligibility
  iTechMedic
  committed
  fc9fb7e
- Merge pull request #135 from halo-box/feat/server-dashboard
  dzannotti
  authored
  f976adb
Commits on Oct 3, 2026
Commits on Oct 4, 2026
- spec : add --spec-n-rs-seq to cap the rollback snapshot slots
  
  Marshall
  authored and
  Laurent Zuijdwijk
  committed
  357ac78
- common : LLAMA_ARG_SPEC_DRAFT_MTP_VOCAB env var for --spec-draft-mtp-vocab
  
  Laurent Zuijdwijk
  and
  claude
  committed
  125dc47
Commits on Oct 6, 2026
- Merge pull request #115 from rcfa/pr/qsa-vulkan-leak
  Nathanw1014
  authored
  e2fd88f
- Merge pull request #119 from rcfa/pr/vulkan-qsa-sel
  Nathanw1014
  authored
  9e22f8c
- hip : mmb per-slice weight decoders for the K-quant and low-bit i-quant types
  
  pwilkin
  and
  claude
  committed
  1b11658
- hip : reshape the routed small tile so every wave issues WMMA
  
  pwilkin
  and
  claude
  committed
  d8cdd7a
- hip : keep fused gate_up experts on MMQ
  
  pwilkin
  and
  claude
  committed
  c17d9de
- hip : admit the measured low-bit expert types to the routed mmb gate
  
  pwilkin
  and
  claude
  committed
  f0945c7
- Merge pull request #141 from michal-zurkowski/fix/mmb-routed-expert-types
  pwilkin
  authored
  ad219f3
- Merge pull request #140 from iTechMedic/fix/hc-l0-gate-f32-output
  dzannotti
  authored
  e3af5e8
- cuda : route small-row wide-batch matmuls to the matvec kernel on AMD
  pwilkin
  committed
  35d5cbd
- CUDA: enable the fused hc_mix_reduce at decode (qwen4exp)
  
  agent
  authored and
  pwilkin
  committed
  389d446
- cuda: IQ3_S vec-dot for the HIP MMVQ path
  
  gaetan-puleo
  authored and
  pwilkin
  committed
  1fabc3a
- CUDA: enable the GDN decode-chain fusion for qwen4exp (rms_norm+scale q/k norm)
  
  agent
  authored and
  pwilkin
  committed
  0d0a722
- CUDA: apply the qwen4exp GDN output gate in the decode norm kernel
  
  gaetan-puleo
  authored and
  pwilkin
  committed
  b912c69
- Merge pull request #78 from michal-zurkowski/pr/pool-pow2
  dzannotti
  authored
  183235a
- Merge pull request #114 from halo-box/fix/rocm-release-lint
  gaetan-puleo
  authored
  4e0d917
- cuda : fuse GDN decode with rollback snapshots
  pwilkin
  committed
  66cfa97
Commits on Oct 7, 2026
- Merge branch 'master' into cuda/small-row-matvec-dispatch
  pwilkin
  authored
  b13befb
- server : add --spec-draft-max-slots to stop drafting under concurrency
  
  dzannotti
  and
  claude
  committed
  3cb54eb
- ci : add on-demand gfx1151 bench and ops lane for the DevLab Dispatch pool
  pwilkin
  committed
  115ed9c
- qsa : track a sequence's prefix while other sequences share the unified cache
  
  rcfa
  and
  claude
  committed
  93f17cd
- llama : load per-layer embedding tables from a separate GGUF (--ple)
  
  michal-zurkowski
  and
  claude
  committed
  4e5b691
- llama-bench : add --ple
  
  michal-zurkowski
  and
  claude
  committed
  c67218b
- qsa : enable the incremental QSA cache on Vulkan
  
  rcfa
  and
  claude
  committed
  24c9276
- gguf-py : add gguf_extract_ple.py to write --ple files
  
  michal-zurkowski
  and
  claude
  committed
  3727eee
- gguf-py : write the PLE tables without faking an ndarray for GGUFWriter
  
  michal-zurkowski
  and
  claude
  committed
  49cd270
- ci : keep the bench steps running when an ops case fails, and report the real fraction
  pwilkin
  committed
  835ed92
- ci : record the first execution in the workflow header
  pwilkin
  committed
  4977288
- Merge pull request #152 from michal-zurkowski/ple-sidecar
  dzannotti
  authored
  c7f11bc
Commits on Oct 8, 2026
Commits on Oct 9, 2026
Commits on Oct 11, 2026