Pull Requests
Community · 3000 Einträge
- [Engine] Inkling-2-Mini from the official NVFP4/MXFP8 checkpoint, expert offload and CPU experts for 8-16 GB cards
- [Engine] add Inkling 2 model family and dual-GPU looped MoE engine support
- [Engine] inkling2 Metal: bind the experts from their own buffer when the model exceeds maxBufferLength
- [Engine] HIP - fix NaN by using per-expert codecs for mixed prefill layers ("the model's output became non-finite (NaN) twice in a row")
- [GUI] control app: native /v1/models rows mirror status, architecture and aliases (#6180)
- [Engine] native Windows support for Intel Arc SYCL, rebased onto 0.2.18 (engine 0.1.64)
- [Engine] q36: --kv fp32|fp16|fp8|fp4 for the K/V cache — 262,144 tokens on one 12 GiB card
- [Engine] session files work with --peer-device
- [Engine] restore a session file without holding its K/V in RAM
- [Engine] inkling2 GGUF loader: read the merged mainline (#32694) layout
- [Engine] gfx906 platinum graft: rp prompt-path stack on v0.2.14.1 — 148 tok/s decode on 4x MI50
- [Engine] inkling2 Metal: let M1/M2 GPUs compile the S3_K/S5_K gate-up kernels
- [Engine] STRATA_VERIFY_QDEDUP: cover the forked shared-expert stream (event-ordered, bitwise-identical image)
- [Engine] sycl inkling: --selfcheck without --native crashed before printing a line
- [Engine] sycl inkling: honor --gpu N — the SYCL backend never read BackendOpts::gpu_index
- [Engine] gfx906: reference build for MI50/MI60/Radeon VII (0.2.11 + int6 kernels)
- [Engine] gfx906 (MI50/MI60/Radeon VII) performance branch — fastest known 4xMI50 Flash-Next numbers
- [Engine] PLE: opt-in NGQ2 n-gram table (2.1 bits a value, 13.4 GB instead of 28.8 GB)
- [Engine] sycl serve: take the hy4/inkling keywords the shared serve sends sampling_keys — every chat through server_intel.py was a 400 since v0.2.18
- [Engine] fix KVA crash with STRATA_GDN_CHUNKED on HIP: the state-only GDN recurrence no longer takes the chunked WMMA scan
- [Benchmark] community: RTX 4090 48 GiB, Qwen3.8-Flash-Next unsloth UD-IQ3_XXS, 262k ctx
- [Engine] sm_70: unrolled Q8_0 GEMV for 2-4 columns (STRATA_Q8_V70, opt-in, bitwise)
- [Engine] sm_70: token-split hyper-connection read (STRATA_GR_V70, opt-in, bitwise)
- [Engine] eliminate sm_86 register and stack spills in strata-inkling CUDA decode kernels
- serve: add opt-in bounded Responses history and continuation
- [Engine] serve: expose context_length/max_context_length on /v1/models entries (#5945)
- [Engine] cpu tests: guard the quota/affinity core count against PR #14's dangling else (#13)
- [GUI] Monitor: an Intel Arc's pkg temperature is the temp sensor, not the pkg energy counter
- [Engine] fix strata-inkling --gpus 2 null deref and VRAM buffer overrun in ep_dual_gpu
- [Engine] support native GGUF experts (Q4_0, IQ4_NL) in MTP drafter
- [Engine] add opt-in zero-spill sm_86 down kernels (STRATA_SM86_DOWN_WS)
- [Engine] Multiple API keys with per-key concurrency limits and reserved capacity
- [TUI] Update the OpenCode providers config to conform to the v2 schema
- [Engine] apply NVFP4 per-expert down scale in grouped subwarp reduction
- [Engine] fix NVFP4 prefill activation quantization past 65,535 CUDA blocks
- [Engine] Merge Ryu's SM x VRAM profile tables into official CUDA Inkling-2
- [Engine] Vision: keep image ownership through request preparation
- [Engine] Pair GDN recurrence inputs and tile prefill kernels in the SYCL engine on Battlemage (BMG)
- [Engine] Group routed projections and reuse expert weights for prefill in the SYCL engine on Battlemage (BMG)
- [Engine] Add opt-in Hy4-Flash expert parallelism for two DGX Sparks
- [Engine] hy4: opt-in HIP build of the multi-GPU expert split
- [Engine] Inkling-2-Mini AMD ROCm/HIP support on top of #6520
- [Engine] Linux: read the experts from the GGUF with O_DIRECT under STRATA_UNBUFFERED_LOAD=1
- [Engine] STRATA_NUMA_ARENA=1 on Windows: node-pinned first-touch slice placement for the expert arena
- [Engine] STRATA_L3_PF=1: +6% decode by reading the next layer's likely CPU experts into the CPU caches while the GPU works (opt-in)
- [Engine] STRATA_ARENA_MMAP on Windows (mapped experts.bin, streamed first fill)
- two-PC: a second PC runs a block of layers as a stage of the layer split (--stage-server / --remote-stage, decode only)
- [Engine] Add opt-in NUMA-aware idle CPU expert prefetch
- [Engine] q36: card-1 expert tier (--peer-device) for Qwen3.6-35B-A3B, 37.9 → 74.8 tok/s
- [Engine] opt-in int8 copies for the persistent hc read (STRATA_HC_PERSIST_Q8, STRATA_HC_I8ONLY)
- [Engine] gfx1030 rocBLAS table up to T=16384, and the ssm_alpha/beta shape (ldy 96) it never matched: 98K/128K prompts -4%
- [Engine] keep the baked qwen think block when preserve_thinking is unset
- [Engine] peer_stream_ring: fix >= 2K peer_i8 ring wraparound + V100 docs
- [Engine] session-identity: the whole STRATA_* environment and the device plan join the config fingerprint (#5379)
- [Engine] session files: format v2 with per-stage save/restore under --layer-split
- [Engine] serve: a tool call written in a shape its schema does not declare is read as the call it meant (#6285)
- [Engine] Opt-in P2P q8_1 decode push, fused P2P down-projection, and GPU peer sync flag
- [Engine] Volta sm_70 decode optimizations, zero-VRAM in-place Q8_0, NVLink split head, and volta-forever detection
- [Engine] Add opt-in streamed-KV QSA slot-group batching, pre75 multi-slot attention, and batch SIDE_MIXER
- [TUI] strata launch --model-path: name Swift's shard 1 as --ple-gguf, where its n-gram table is
- [Engine] Add opt-in multi-slot batched MTP drafting (STRATA_BATCH_MTP_FUSED)
- [Engine] Opt-in resize_for_overwrite and parallel host memcpy for kv_mode != 0 park/restore
- [Engine] Size STRATA_PARK_ASYNC VRAM stage by device-resident bytes under kv_mode != 0
- [Engine] Fix O(N^2) scan in getline_fd and avoid full-line copy in BGEN/parse_i64_list
- [Engine] Guard negative page_table entries in qsa_prompt_attn and track monotonic kv_stream overflow_count
- [GUI] Keep quiet desktop chat streams alive
- Two cards of different makes behind one server
- [Engine] SYCL: enable the conversation cache's SSD tier (0.2.17's ConversationSpill)
- [Engine] Persist the SSD conversation tier across restarts when the model fingerprint matches
- [Engine] fingerprint the resolved MTP files, so the SSD conversation tier stays on for GGUF and rt drafters
- [Engine] Hy4-Flash: SQ-96 starts on 128 GB Macs, --info no longer crashes, session reuse on by default, reasoning_effort max in configs
- [Engine] Hy4-Flash (Metal): resume at the longest common prefix, several conversation slots, SSD tier kept across restarts
- [Engine] SSD conversation tier: fingerprint the resolved MTP files and keep the tier across restarts (v0.2.18)
- [Engine] sycl: keep the branch point a follow-up can reach
- [Engine] session-file saves: ping-pong blocks and a writer thread so the source copy overlaps the previous block write