Commit Graph

  • 146df3c468 setup.py run(): the app's installer also runs the bundled Mac preparation tools in metal/tools (metal_repack.py, stream_experts.py); a fresh Mac Terminal or app install of UD-Q4_K_XL or UD-IQ4_XS-v31 stopped with "only runs its bundled preparation tools" since af1ffd913. Test: they run, and -m, another interpreter and other folders are still refused main v0.2.12 Niko1221 2026-11-07 01:29:36 +01:00
  • 9780a8c27d Merge main (54828c7ef): four sponsors in Loyal supporters, all 8 READMEs Niko1221 2026-11-06 21:53:35 +01:00
  • 311926416a Merge f-0212-tools (ea765a571): tool-call fixes in serve/ and the native gateway, with their tests and docs Niko1221 2026-11-06 21:53:34 +01:00
  • 0fe4ae5090 qsa_select tp_select: no __match_any_sync below sm_70 (a plain add per thread there; integer atomics, the same sums): the CUDA 12 release engine (sm_60..89) builds again Niko1221 2026-11-06 21:39:07 +01:00
  • ea765a5717 App gateway: an unknown model name on /v1/messages uses the one loaded model (Claude Desktop sends claude-5-5-sonnet); chat and Responses stay exact, and with several models loaded the 400 stays (#5312) Niko1221 2026-11-06 19:55:58 +01:00
  • c49f198c77 Test: the Anthropic route serves the loaded model whatever name the request carries (#5312, claude-5-5-sonnet) Niko1221 2026-11-06 19:51:51 +01:00
  • b1541ab245 Parser: the prefixed-name and misspelt-opener checks look only at the end of the buffer (the new text and 256 characters before it), so a long call is not rescanned with every token (#5409, #5276) Niko1221 2026-11-06 19:48:21 +01:00
  • 8a42a9457d moe_iq4_w4a8: the device side compiles for sm_60..75 targets (a trap there, as moe_fused_iq.cu does; arch_ok() already declines those cards, so no behaviour changes): the CUDA 12 and CUDA 13 release engines build again Niko1221 2026-11-06 19:42:58 +01:00
  • 05506f9532 Tool names: a namespaced MCP tool written by its last segment (get_asset_info for mcp__assets__get_asset_info) is that tool when exactly one declared tool ends that way (#5281) Niko1221 2026-11-06 19:40:40 +01:00
  • 95381770f8 Anthropic tool results follow tool_use_id: results sent in the order the tools finished are put in call order before the template pairs them by position (PR #4832) Gabriel A. Devenyi 2026-11-06 19:36:59 +01:00
  • d7e555bdbb tool_choice "required" with several tools: write only "<tool_call>\n", because from a written "<function=" the model went on with names no tool has and the reply came back as text (PR #4828) Gabriel A. Devenyi 2026-11-06 19:36:43 +01:00
  • 75acfca321 Config undeclared_tool_calls (env STRATA_UNDECLARED_TOOL_CALLS): "call" returns a well-formed call of a tool the request did not declare as a tool call; the default "text" is as before (#4622) Niko1221 2026-11-06 19:36:04 +01:00
  • 63fd759227 Tool calls: a misspelt opener (<toolcall>, <tool-call>) followed by <function=NAME> of a declared tool opens the call, where a <tool_call> would; prose and code stay text (#5276) Niko1221 2026-11-06 19:34:47 +01:00
  • 54828c7efb README: Loyal supporters adds themactep, macintog, bisonbet and vizv (public GitHub sponsors) in all 8 languages Niko1221 2026-11-06 19:33:46 +01:00
  • 68ab15ddc1 Tool names: <function=tools NAME> (a space for the dot) is the declared tool NAME, and a bare <function=tools> is the one declared tool that takes the call's parameters (with tool_call_recovery); anything else stays text (#5409) Niko1221 2026-11-06 19:32:59 +01:00
  • 7be42963e6 Version 0.2.12 / engine 0.1.58: Terminal and app 0.2.12, Control and every engine package 0.1.58 (the 0.2.11 bump list, plus the v0.1.58 line in the setup.py version history) Niko1221 2026-11-06 19:21:19 +01:00
  • 5b1728358b test_download_integrity: a server that keeps serving the wrong file ends with the file kept as .bad (the dlsha behaviour: the new mismatch line, the file not deleted) Niko1221 2026-11-06 17:43:19 +01:00
  • 706efc5b0f prefill_fused_iq_test: skip the Q5_0 down pairs where the native kernels do not cover them (CUDA has none; the gfx12 / gfx1151 builds run them); --only=NAME still insists Niko1221 2026-11-06 16:29:26 +01:00
  • 0a0e4463ea file tier: the re-check after the RAM copy counts only the page cache the group can really have (on by default; STRATA_FILE_TIER_HONEST=0 turns it off) Niko1221 2026-11-06 15:50:23 +01:00
  • e77bd43bc9 prefill: a Q5_0 down (UD-IQ4_XS-v31: 33 of 48 layers) on the fused int8 kernel on gfx12, opt-in STRATA_PF_FUSED_Q50=1 Niko1221 2026-11-06 15:32:06 +01:00
  • 37b5cf0db2 gfx12: the start-up line for the length-gated chunked GDN scan, and STRATA_GFX12_DEFAULTS=0 turns the gate off Niko1221 2026-11-06 15:31:23 +01:00
  • c0e7a6f4f8 gfx12: the chunked GDN scan is the default for prompt chunks of 30,000 tokens or more Niko1221 2026-11-06 15:23:52 +01:00
  • 645b9177e3 setup: wording for the kept .bad file (NEEDS OWNER APPROVAL): the mismatch line says the file was kept, and says when the server sends the very same bytes; one line when the kept file is removed for room Niko1221 2026-11-06 13:37:29 +01:00
  • 071e4dc51e setup: a downloaded file with the wrong SHA-256 is kept as <name>.bad instead of deleted; the next run hashes it again and takes it back if it matches, a good new download or too little room removes it (visible text unchanged in this commit) Niko1221 2026-11-06 13:37:11 +01:00
  • bca7931f3d peer tier: STRATA_ADAPT_OVERLAP=1 with --peer-device (the peer's copies land while the next window runs) bulhenvaster 2026-11-06 08:51:15 +01:00
  • 2b9f05093d prefill: STRATA_PF_PEER_CPU=1 (opt-in) runs the CPU share beside a peer GPU bulhenvaster 2026-11-06 08:48:03 +01:00
  • 24c6c2f682 Admit the PCIe share's fetched misses to the GPU cache (opt-in STRATA_ADAPT_FETCH=1|2) sergqwer 2026-11-06 15:30:25 +01:00
  • 386f9959f1 Exact two-pass QSA prompt selection (opt-in STRATA_SELECT_TWO_PASS=1) bitaccesscomau 2026-11-06 10:26:21 +01:00
  • f281e3fe26 force_check.py: absolute path for the profile copy Niko1221 2026-11-06 10:49:08 +01:00
  • f46e88d610 force_check.py: cfg profile_src gives each engine start a fresh copy of the expert profile Niko1221 2026-11-06 10:48:58 +01:00
  • d486fbe7d6 serve: STRATA_FOLLOW / STRATA_FOLLOW_LOGPOS token forcing through the real decode/verify windows (test and measurement only, off by default); force_check.py score --path decode Niko1221 2026-11-06 10:38:43 +01:00
  • 65909b2d0f verify: STRATA_VERIFY_PROFILE and STRATA_ROUTE_STATS reports cover windows up to 16 rows (measurement only; default behaviour unchanged); tools/force_check.py: the token-forcing check (scorer driver + the paper's two rules) Niko1221 2026-11-06 10:35:53 +01:00
  • cae562fa01 verify: STRATA_FORCE_FAULT=<m>:<bias> test knob for the token-forcing scorer (adds a bias to every m-th token's logit in the STRATA_LOGPOS rows; off by default, silent) Niko1221 2026-11-06 10:29:06 +01:00
  • 63f063852f prefill: UD-IQ4_XS-v31's Q5_0 down layers (33 of 48, with IQ3_XXS or IQ4_XS gate/up) take the fused F16 matrix-core kernel on gfx1151 under Windows instead of leaving the fused path for gather + FP16 dequantize + hipBLASLt. The Q5_0 block decode (qh bits spread, int8 codes q - 16, scale d) is the RDNA4 patch's; native_f16_kernel and the FP16 down rows are the K-quant experts' route. On by default there; STRATA_PF_FUSED_Q50=0 or STRATA_GFX1151_DEFAULTS=0 turns it off (the layers keep the earlier route), =1 turns it on for any gfx1151 build; it needs the F16 prompt set (STRATA_PF_F16). Other GPUs are untouched. Parity test pairs IQ3_XXS / Q5_0 and IQ4_XS / Q5_0 against the FP32 dequant reference, with 256/64, 33/16, 1/16, 130/32 and 517/32 token/expert cases: rel RMS 2.7e-3 against the int8-rounded model, 1.4-1.5e-2 against FP32 (the IQ pairs' usual level) Niko1221 2026-11-06 15:22:53 +01:00
  • cc3f9142e3 Auto on the 12 GB / streaming CUDA class: IQ3_S starts the MTP, UD-IQ4_XS-v31 starts DFlash2 with the tuned pack at the class's own 0.8 Niko1221 2026-11-06 15:37:13 +01:00
  • f138653938 DFlash2 on Strix Halo: the tuned pack for UD-IQ4_XS-v31 too, at Strix's current confidence Niko1221 2026-11-06 12:04:46 +01:00
  • ce6bbd7d54 DFlash2 on Intel Arc: the tuned bf16-selector pack for SYCL + IQ3_XXS at Arc's own confidence 0.6 Niko1221 2026-11-06 11:37:39 +01:00
  • aba4ddaca7 Auto on HIP gfx1201 starts the MTP until the DFlash2 prompt stall is fixed (one entry: drafting.AUTO_MTP) Niko1221 2026-11-06 10:56:13 +01:00
  • 31928fbe34 DFlash2: Strata's tuned drafter pack on CUDA (every expert resident) and Strix Halo, at their current confidence Niko1221 2026-11-06 10:40:01 +01:00
  • 10831eceb2 Metal build: stubs for the f-0216-cpu 512-bit entry points (q2_rows_512_ok, iq512f_supported, iq512_fused_gu_rows, iq512_reference_gu_rows, q50_512_down_rows) Niko1221 2026-11-06 09:57:08 +01:00
  • 7604725595 arm64 portable build: stubs for cpu_vnni_ok, q2_rows_512_ok and iq512_reference_gu_rows (the Spark link failed on them) Niko1221 2026-11-06 09:51:02 +01:00
  • 2a8217c0ce generate: the CPU expert kernel line reads 'CPU expert kernels: <X>' (owner's short wording) Niko1221 2026-11-06 09:44:37 +01:00
  • a7cc8faa5d split: one constant for the expert window floor, and the IQ3_S pool test Niko1221 2026-11-06 06:28:23 +01:00
  • 20885f7af0 split: retain a window in every expert pool MIDHAT NASHAR 2026-11-06 06:25:15 +01:00
  • abf2953be0 native_dense: STRATA_HC_Q8 copies only the stage's own Q8_0 hyper-connection rows on a trimmed layer split (P1.2) Niko1221andhavok2-htwo 2026-11-06 06:25:10 +01:00
  • 78c5a834e3 Merge f-0216-dfship into f-0216-all Niko1221 2026-11-06 09:36:34 +01:00
  • 51548e5539 Merge f-0216-strix2 into f-0216-all Niko1221 2026-11-06 09:36:34 +01:00
  • 7597e383cf Merge f-0216-rtx into f-0216-all Niko1221 2026-11-06 09:36:30 +01:00
  • f2eedfe9f2 Merge f-0216-cpu into f-0216-all Niko1221 2026-11-06 09:36:30 +01:00
  • bd0e878f42 Merge f-0216-core into f-0216-all Niko1221 2026-11-06 09:36:29 +01:00
  • 4084979c95 prefill: STRATA_PF_ALLOC_TRACE=1 prints the free-memory figure around Prefill::init's allocations against what the code asked for (stager ring, routing tables, PLE rows, GEMM scratch, the chunk's buffers, library warm-up). Off by default, no behaviour change Niko1221 2026-11-06 08:56:14 +01:00
  • 530a29ea2d serve: on an integrated GPU with a page-locked expert complement (Strix Halo), the free-VRAM startup line and the request-start warning no longer advise a larger --vram-reserve-mib (it cannot raise that figure: the driver counts pinned host memory and the prompt path's buffers about twice, so the cache giving up experts to the complement only lowers it). New startup line: "N MiB of graphics memory reported free. On this PC the graphics chip shares the system memory, so this figure reads low, and a larger --vram-reserve-mib won't raise it. If decode drops to a few tok/s, lower --max-context." New request-start note, printed once: "Only N MiB of graphics memory reported free at the start of this request. On this PC the figure reads low because the graphics chip shares the system memory. If decode drops to a few tok/s, lower --max-context." Every other GPU keeps the earlier text. Wording only: greedy ids identical to engI on agent8k, agent, pp4k and code_py (UD-Q4_K_XL, setup config) Niko1221 2026-11-06 08:56:13 +01:00
  • 3e8d3b2940 Mac: DFlash2 is the default drafter for UD-Q4_K_XL and UD-IQ4_XS-v31, with Strata's tuned drafter at --draft-conf 0.7 Niko1221 2026-11-06 07:52:31 +01:00
  • 039cecc5a5 feed: a drafters member for DFlash2 drafter files Strata hosts (pack-drafter, build-feed --drafters, make_local_feed --drafter) Niko1221 2026-11-06 07:52:31 +01:00
  • d735d66a4d docs: STRATA_IQ4_W4A8 on sm_120 and its exit counters; STRATA_FILE_TIER_IDLE and the idle policy of the file tier Niko1221 2026-11-06 01:49:09 +01:00
  • 7bf5ec4ebf prefill: opt-in int8 kernel for the IQ4_XS / IQ4_NL expert products (STRATA_IQ4_W4A8), MMQ's bits cctan 2026-11-06 01:49:09 +01:00
  • 178d1c74b5 merge f-0215-mac-pp (host-row pipeline, qsa_merge_h, ragged expert tiles) into f-0215-mac-int Niko1221 2026-11-06 01:25:09 +01:00
  • eecd7f61d7 merge f-0215-mac-out (step profile, opt-in screens, Metal DFlash2 conf 0.5, calibrated gate) into f-0215-mac-int Niko1221 2026-11-06 01:25:09 +01:00
  • 86c3367a89 native_expert.cpp: the fused-rows type list terminator is '\0' (a NUL byte had slipped into the character constant; MSVC rejects it) Niko1221 2026-11-06 01:05:40 +01:00
  • 04b73d061d kq_window_bench: --dump-trace writes the synthetic windows in the route-dump text form Niko1221 2026-11-06 01:01:14 +01:00
  • 1daec2a5b8 check_isa_guard: iq512_reference_gu_rows and iq512x_rows are listed (main already reports them as unlisted; both are reached only behind cpu_avx512_ok / cpu_vnni_ok) Niko1221 2026-11-06 00:51:10 +01:00
  • b94ab5f4c6 kq_window_bench: the CPU pool's part of a decode window replayed without a GPU (per-layer format pairs from the pack, expert and row counts calibrated on the engine's logged means, cold DRAM) Niko1221 2026-11-06 00:50:40 +01:00
  • 445cdf4540 Run manifest: say what the unset STRATA_ROUTE_TAIL_SKIP means (the engine default applies, tails are skipped only where the configuration allows) Niko1221 2026-11-06 00:47:28 +01:00
  • 8463d15c84 STRATA_ROUTE_DUMP: terminator fix Niko1221 2026-11-06 00:42:08 +01:00
  • 1c768f9d12 STRATA_ROUTE_DUMP: newline fix Niko1221 2026-11-06 00:41:18 +01:00
  • b0b0348c2b kernels: opt-in fused gate/up rows in ggml order for 2+ tokens (STRATA_IQ512_FUSED_MAX; N, after PR 5243's row_gu_dot); STRATA_ROUTE_DUMP routing trace for offline placement replays Niko1221 2026-11-06 00:39:11 +01:00
  • 7e3fae57ca PR 5120 tier below VBMI is opt-in (STRATA_VNNI_TIER=1): its 512-bit rows round differently from the AVX-2 rows such a CPU runs today (SDE Cascade Lake: all multi-token gate/up lines and the Q2_0 rows differ), so by default every CPU keeps its output Niko1221 2026-11-06 00:22:37 +01:00
  • 32944ce1ee cpu_dispatch_dump: the Q2_0 rows (q2_rows_any) next to ggml's, so the substitute unpack can be compared with the multishift Niko1221 2026-11-06 00:20:07 +01:00
  • f000d4a455 cpu/experts: the 512-bit VNNI rows on a CPU without AVX-512-VBMI (PR 5120) Michel Pollet 2026-11-06 00:20:07 +01:00
  • 6368dc68ab cpu: drop the IQ4_NL 512-bit down rows (5-11% slower than iq4nl256_down_rows at 2-4 tokens on Zen 4); Q5_0 down keeps its 512-bit rows; vpdpwssd in the i-quant rows' integer sums (same integers, one op less) Niko1221 2026-11-06 00:19:58 +01:00
  • 55dc17697c cpu: IQ4_XS gate/up and Q5_0 / IQ4_NL down rows in 512-bit lanes on AVX-512 CPUs, bit-identical to the code they replace Niko1221andbsorensen110 2026-11-06 00:02:58 +01:00
  • 241b2f203c cpu_dispatch_dump: a hash per (pair, token count, row range) of the pool's native expert calls, to compare a change switched off and on Niko1221 2026-11-06 00:02:58 +01:00
  • 044105d8de kq_cpu_bench: --part gu|down times one projection alone, --hot K walks K distinct experts (cache-resident case) next to the DRAM walk Niko1221 2026-11-05 23:20:48 +01:00
  • 994029b02f generate: the file tier's Linux defaults and the routing-aware prefetch stay off while every expert is in the GPU caches Phlimosx 2026-11-06 00:22:04 +01:00
  • 89f4fa18ee Run manifest (lite): STRATA_MANIFEST=<file> writes the selected fast paths, fallback reasons, tensor formats, engine choice and budgets as JSON (P0.1) Niko1221 2026-11-06 00:21:53 +01:00
  • e2d3159996 Build the flag-order test on HIP too: hip_flag_order_test is the same source through the compat header (P1.1) Niko1221andAlanJager 2026-11-06 00:20:25 +01:00
  • 1ed923f59e PLE reader: issue() folds a small batch in a stack buffer instead of a hash table and a job vector (P1.6) Niko1221andmartifaj 2026-11-06 00:12:28 +01:00
  • aeaca205eb Mixed-HIP pipelined prompt read: stage 0 is served to its end before the drafter's blocking prefill on its card (P1.9, part 2) Niko1221andcloud4t0r 2026-11-05 23:57:46 +01:00
  • f225844d19 Prompt MMQ row tables get the stream-k fixup's over-read margin, sized from the backend's tile bound (P1.9, part 1) Niko1221andquantforge-br 2026-11-05 23:46:57 +01:00
  • 40849df0a8 Transfer completion: flag B stays behind queued copies, and the adaptive-copy range covers the whole round (P1.1) 2026-11-05 23:39:06 +01:00
  • a5b820a415 Request accounting and prompt timing marks: yielded prompt legs summed once, admission time without a double count, unknown stays unknown (P0.4) 2026-11-05 23:31:14 +01:00
  • b97433df05 spec: STRATA_DF2_GATE=1 (opt-in): DFlash2's verify window from a per-request calibrated survival (position x confidence bucket, prior C^1.3) against the cost of one more row, instead of a fixed --draft-conf; window_rule_test covers it Niko1221 2026-11-05 23:18:51 +01:00
  • 64cc20a8e2 Merge f-0215-fit (setup wording) into f-0215-int Niko1221 2026-11-05 22:47:02 +01:00
  • 1d342f0836 setup: on a unified-memory GPU whose RAM budget leaves no room for a numeric expert cache, say where the cache is sized from instead of warning that it may take the OS's room: "The expert cache is sized from the graphics memory (<N> GB carve-out + shared), not from the RAM." N is the carve-out setup detected (no number when none was). The old warning and its --resident-budget-gib tip are gone; the test checks the new line with 2 GB and 48 GB carve-outs Niko1221 2026-11-05 22:44:44 +01:00
  • 0e1d440713 Merge f-0215-pk into f-0215-int Niko1221 2026-11-05 22:36:02 +01:00
  • 19136808f9 Merge f-0215-fit into f-0215-int Niko1221 2026-11-05 22:35:59 +01:00
  • 2c0ad33f71 Metal prompt: ragged expert tiles on by default for the types that have kernels (STRATA_PROMPT_V6=0 off); drop the unused Mac PLE threading and the tall-tile probes Niko1221 2026-11-05 21:59:37 +01:00
  • 3cc5714f37 Strix Halo (Windows gfx1151): STRATA_SMALL_UNR=3 and STRATA_GR_Q8_PFS=1 (load-ordering variants of the decode kernels, same arithmetic) join the default table. UD-Q4_K_XL, DFlash2 serve, 4 rounds x 2 passes x 4 prompts interleaved: window time -0.25% (25 of 32 pairs faster, same window and token count in every pair); first-token logits and 220-token outputs identical on 3 prompts for UD-Q4_K_XL and UD-IQ4_XS. Linux gfx1151 and STRATA_DEFAULTS=026 keep their tables; arch_defaults_test updated and passing Niko1221 2026-11-05 21:19:24 +01:00
  • 687a7d79d7 Metal prompt: host rows beside the GPU (embedding + PLE rows of the next chunk on a thread, parallel PLE gather on Macs), qsa_merge_h (merge writes the fp16 operand), ragged expert tiles (mm_gu6/mm_down6/mm_scan6, STRATA_PROMPT_V6), STRATA_PROMPT_KPROF per-dispatch profile, matmul2d probes Niko1221 2026-11-05 20:36:54 +01:00
  • 6e31946cbe Strix Halo (gfx1151, Windows): the F16 prompt expert kernels write the down rows as FP16 and a combine kernel reads them (STRATA_PF_DM16, on by default there, =0 turns it off; from the Strix dm16 work patch, extended to the F16 K-quant and IQ kernels). 8K prompt, same engine, interleaved pairs: IQ3_S +2.2% (4 of 4), UD-IQ4_XS +2.6% (3 of 3), UD-Q4_K_XL +1.2% (3 of 4); first-token KL 1e-4 at 8K, 0.02-0.04 on agent transcripts Niko1221 2026-11-05 19:59:50 +01:00
  • 66b346c29f drafting: Metal gets its own DFlash2 confidence (0.5, BACKEND_CONF) instead of the CUDA slot steps, which wrote 0.8 for the Mac's --expert-cache 64; M5 Max UD-Q4_K_XL +14.3% over MTP spec 2 on six kinds Niko1221 2026-11-05 19:54:31 +01:00
  • 830732a33d metal: step profile scales every pass by the cool whole-window time; opt-in screens STRATA_METAL_HCLD=1 (HC loads in flight: -10.9%, off), STRATA_METAL_WPF_MIB (weight prefetch beside gr_up), STRATA_METAL_MOE_GU / MOE_DN (expert tile override) Niko1221 2026-11-05 19:54:31 +01:00
  • e431a4a8e2 Strix Halo (Windows gfx1151): the V2K grouped decode experts for Q4_K / Q5_K gate/up and Q5_1 / Q8_0 down are on by default (STRATA_EXPERT_V2K=0 turns them off). UD-Q4_K_XL, DFlash2 serve, 16 request pairs: window time -4.0% (16 of 16 faster, mean x1.040), same window and token count in every pair; first-token logits and 220-token outputs identical on 3 prompts (the kernels are bitwise the old ones) Niko1221 2026-11-05 19:27:35 +01:00
  • 4dbaf703fd Strix Halo (gfx1151, Windows): the K-quant prompt experts of UD-Q4_K_XL (Q4_K / Q5_K gate/up, Q5_1 / Q8_0 down) take the fused F16 matrix-core kernels by default (STRATA_PF_FUSED_KQ=0 or STRATA_GFX1151_DEFAULTS=0 keeps the earlier route). Setup config, n=1: 8K prompt 885 -> 1,387 tok/s, 4K 611 -> 1,316; first-token KL against the earlier route 1e-5 .. 6e-2 on prompts of 1K+ tokens (information), same greedy ids Niko1221 2026-11-05 18:54:42 +01:00
  • c4fa1d9e1e metal: STRATA_METAL_STEPPROF=T, the in-situ step profile by part (prefix runs at every barrier group, weight bytes and GB/s per part, distinct experts per layer); run_graph can run the first N dispatches Niko1221 2026-11-05 18:41:45 +01:00
  • 9f7d977d22 README: a Loyal supporters section under the Sponsor and coffee buttons in all 8 languages, with the public GitHub sponsors as linked profile pictures (first: ryan4yin) and a one-line template for the next one Niko1221 2026-11-05 14:54:54 +01:00
  • 5d37280db1 Docker: build.sh copies metal/tools/metal_repack.py and stream_experts.py into the build context (the Terminal bundle carries them since the Mac Unsloth sets; the v0.2.11 image build stopped at the missing metal_repack.py); the context test checks every INSTALLER_METAL_TOOLS name Niko1221 2026-11-05 14:02:27 +01:00
  • dc9dfcfd29 GitHub Sponsors: FUNDING.yml names github: Niko1221 (the repository's Sponsor button), and every README gets a Sponsor badge beside the coffee badge, a Sponsor button in Support Strata and a support line that names both Niko1221 2026-11-05 14:02:27 +01:00
  • c3b3f50355 Merge f-0214-int d7f5c6dd1 (Strix first-launch plan fix: expert cache auto on UMA, Python only) into f-0211-rc v0.2.11 Niko1221 2026-11-05 12:54:06 +01:00
  • d7f5c6dd11 Merge f-0214-strixplan 0c0b7b875 (Strix plan fix) into f-0214-int Niko1221 2026-11-05 12:53:44 +01:00