Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
dsv41: the CUDA build of the tree needs two portable edits (found on z440's sm_86 bring-up)
Niko122145dc5e340085
diff --git a/CMakeLists.txt b/CMakeLists.txtindex e07c2462f..76a8801f1 100644--- a/CMakeLists.txt+++ b/CMakeLists.txt@@ -797,7 +797,15 @@ if(STRATA_ENABLE_CUDA OR STRATA_ENABLE_HIP)endif()if(STRATA_NATIVE_EXPERTS)# the trellis experts (device kernels ds4_tq.cu, CPU lane ds4_tq_cpu.cpp) against a double reference; needs a GPU, no model- set_source_files_properties(tests/dsv41/tq_expert_test.cpp PROPERTIES LANGUAGE HIP)+ # both test .cpp files launch kernels, so they compile as HIP on the AMD backend and as CUDA on the NVIDIA one+ if(STRATA_ENABLE_HIP)+ set(_ds4_test_lang HIP)+ elseif(STRATA_ENABLE_CUDA)+ set(_ds4_test_lang CUDA)+ else()+ set(_ds4_test_lang CXX)+ endif()+ set_source_files_properties(tests/dsv41/tq_expert_test.cpp PROPERTIES LANGUAGE ${_ds4_test_lang})add_executable(ds4_tq_test tests/dsv41/tq_expert_test.cpp src/dsv41/ds4_tq.cu src/dsv41/ds4_iq.cu src/dsv41/ds4_tq_cpu.cppsrc/dsv41/ds4_iq_cpu.cpp src/dsv41/ds4_gg_cpu.cpp tools/trellis/tqenc.cpp)target_include_directories(ds4_tq_test PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/include ${_strata_gpu_include_directories}@@ -819,7 +827,7 @@ if(STRATA_ENABLE_CUDA OR STRATA_ENABLE_HIP)endif()# the i-quant experts (ds4_iq.cu) against ggml's own decode of the same blocks + a double reference, device and CPU# lane, for every type the device takes: the gate a new type (IQ2_XS ..) has to pass before the engine uses it- set_source_files_properties(tests/dsv41/iq_expert_test.cpp PROPERTIES LANGUAGE HIP)+ set_source_files_properties(tests/dsv41/iq_expert_test.cpp PROPERTIES LANGUAGE ${_ds4_test_lang})add_executable(ds4_iq_test tests/dsv41/iq_expert_test.cpp src/dsv41/ds4_iq.cu src/dsv41/ds4_iq_cpu.cppsrc/dsv41/ds4_gg_cpu.cpp src/dsv41/ds4_tq.cu src/dsv41/ds4_tq_cpu.cpp)target_include_directories(ds4_iq_test PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/include ${_strata_gpu_include_directories}diff --git a/src/dsv41/ds4_tq.cu b/src/dsv41/ds4_tq.cuindex 04b994d40..aab0a76d7 100644--- a/src/dsv41/ds4_tq.cu+++ b/src/dsv41/ds4_tq.cu@@ -259,7 +259,7 @@ __device__ __forceinline__ void tile_dot(const uint8_t* const* wrows, int nblk,#pragma unrollfor (int i = 0; i < 8; ++i) {const float p0 = part[mi][i], p1 = part[mi][8 + i];- const float q0 = __shfl_xor(p0, 16), q1 = __shfl_xor(p1, 16);+ const float q0 = __shfl_xor_sync(0xffffffffu, p0, 16), q1 = __shfl_xor_sync(0xffffffffu, p1, 16);tot[mi][i] += d[mi] * (kg == 0 ? p0 + q0 : p1 + q1);}#endifdiff --git a/tools/dsv41/ista_intake.py b/tools/dsv41/ista_intake.pynew file mode 100644index 000000000..6626e8aa1--- /dev/null+++ b/tools/dsv41/ista_intake.py@@ -0,0 +1,291 @@+#!/usr/bin/env python3+"""ista_intake.py - one command per IST-DAS commit (DEEPSEEK-ISTA-PLAN Phase 0).++For a pinned HF commit and one or more of its sets it:+ 1. lists the repo tree at that commit (path + size) and writes it to the report folder;+ 2. downloads only files that are absent or have the wrong size, into <root>/<set>/;+ 3. hashes everything into <root>/SHA256SUMS-<commit>-<set>.txt and hard-links duplicates (the Engram file is the same+ 110.6 GB blob in all three sets, so it is stored once);+ 4. reads the GGUF headers and reports every tensor type with its count and bytes, flagging types the device kernels do+ not take (the same set --pack-info prints as "no GPU kernel");+ 5. builds a pointer pack with pack_gguf.py - no requant, the file as published;+ 6. runs --pack-info, the scripted eval (session16.sh, 4995 positions) and a speed smoke (code + agent, DSpark off/on);+ 7. writes one report per commit and set: <report-dir>/<commit>/<set>.md, every number tagged with the commit.++Nothing here requantises or rewrites a GGUF. Exit codes: 0 ok, 1 a stage failed (the report says which), 2 bad usage.++usage:+ ista_intake.py --commit 9e2528f --sets 2.00bpw,2.25bpw+ ista_intake.py --commit 9e2528f --sets 2.00bpw --no-eval # download + pack only+ ista_intake.py --commit 9e2528f --sets 2.00bpw --no-download # headers + pack + eval on what is on disk+"""+from __future__ import annotations++import argparse+import hashlib+import json+import os+import pathlib+import re+import shutil+import subprocess+import sys+import time++REPO = "helcig/tmp-oct26"+ROOT = pathlib.Path("/data/helcig")+REPORTS = pathlib.Path("/data/dsv41-port/runs/intake")+PORT = pathlib.Path("/data/dsv41-port")+PY = "/data/dsv41/hfvenv/bin/python"+HF = "/data/dsv41/hfvenv/bin/hf"+TOOLS = PORT / "src/tools/dsv41"+# the substitutions the published sets need today (HANDOVER.md §6; Phase 1 removes them)+ENGRAM_DIR = pathlib.Path("/home/niko/engram-ista/2.00bpw")+DSPARK = pathlib.Path("/data/dsv41/source/dspark-DeepSeek-V4.1-Flash-Q8_0.gguf")+DENSE_WOA = pathlib.Path("/home/niko/dsv41-dense/wo_a_q5k")+DENSE_HEAD = pathlib.Path("/home/niko/dsv41-dense/head_q6k")+# ggml type ids the device kernels take (src/dsv41/ds4_iq.cu's supported() plus MXFP4); kept in sync with --pack-info+DEVICE_TYPES = {0, 1, 2, 3, 8, 10, 11, 12, 13, 14, 16, 17, 18, 19, 20, 21, 22, 23, 29, 39}+TYPE_NAMES = {0: "F32", 1: "F16", 2: "Q4_0", 3: "Q4_1", 6: "Q5_0", 7: "Q5_1", 8: "Q8_0", 9: "Q8_1",+ 10: "Q2_K", 11: "Q3_K", 12: "Q4_K", 13: "Q5_K", 14: "Q6_K", 15: "Q8_K",+ 16: "IQ2_XXS", 17: "IQ2_XS", 18: "IQ3_XXS", 19: "IQ1_S", 20: "IQ4_NL", 21: "IQ3_S",+ 22: "IQ2_S", 23: "IQ4_XS", 24: "I8", 25: "I16", 26: "I32", 27: "I64", 28: "F64",+ 29: "IQ1_M", 30: "BF16", 39: "MXFP4"}+++def log(msg: str) -> None:+ print("%s %s" % (time.strftime("%H:%M:%S"), msg), flush=True)+++def run(cmd: list[str], **kw) -> subprocess.CompletedProcess:+ return subprocess.run(cmd, capture_output=True, text=True, **kw)+++def sha256_file(p: pathlib.Path, chunk: int = 1 << 24) -> str:+ h = hashlib.sha256()+ with open(p, "rb") as f:+ while True:+ b = f.read(chunk)+ if not b:+ break+ h.update(b)+ return h.hexdigest()+++def repo_tree(commit: str) -> list[tuple[str, int]]:+ from huggingface_hub import HfApi+ rows = []+ for f in HfApi().list_repo_tree(REPO, revision=commit, recursive=True):+ size = getattr(f, "size", None)+ if size:+ rows.append((f.path, int(size)))+ return sorted(rows)+++def local_path(root: pathlib.Path, repo_path: str) -> pathlib.Path:+ return root / repo_path+++def download(commit: str, repo_path: str, dest_root: pathlib.Path) -> int:+ dest_root.mkdir(parents=True, exist_ok=True)+ r = run([HF, "download", REPO, "--revision", commit, "--include", repo_path, "--local-dir", str(dest_root)])+ if r.returncode != 0:+ log(" download failed: %s" % (r.stderr or r.stdout)[-300:])+ return r.returncode+++def tensor_types(shards: list[pathlib.Path]) -> dict[int, tuple[int, int]]:+ """type id -> (tensor count, bytes) over the shards, using pack_gguf's Gguf reader and its block table."""+ sys.path.insert(0, str(TOOLS))+ import pack_gguf # noqa: E402+ types = dict(pack_gguf.TYPES)+ # pack_gguf's table covers the quant types; the intake also meets these (ggml ids, values/block, bytes/block)+ types.update({2: ("Q4_0", 32, 18), 3: ("Q4_1", 32, 20), 6: ("Q5_0", 32, 22), 7: ("Q5_1", 32, 24),+ 9: ("Q8_1", 32, 36), 15: ("Q8_K", 256, 292), 19: ("IQ1_S", 256, 50), 29: ("IQ1_M", 256, 56),+ 24: ("I8", 1, 1), 25: ("I16", 1, 2), 26: ("I32", 1, 4), 27: ("I64", 1, 8), 28: ("F64", 1, 8)})+ out: dict[int, tuple[int, int]] = {}+ for s in shards:+ g = pack_gguf.Gguf(s)+ for name, typ, dims, off in g.tensors: # (name, type, dims, offset); the file handle is closed already+ t = types.get(typ)+ if t is None:+ out[typ] = (out.get(typ, (0, 0))[0] + 1, out.get(typ, (0, 0))[1])+ continue+ nvals = 1+ for d in dims:+ nvals *= int(d)+ nbytes = (nvals // t[1]) * t[2] if t[1] > 1 else nvals * t[2]+ c, b = out.get(typ, (0, 0))+ out[typ] = (c + 1, b + nbytes)+ return out+++def pack_set(commit: str, set_name: str, root: pathlib.Path, report: list[str]) -> pathlib.Path | None:+ setdir = root / set_name+ pack = PORT / "packs" / ("ista_%s_%s" % (set_name, commit[:7]))+ if pack.exists():+ report.append("- pack already exists: %s" % pack)+ return pack+ cmd = [PY, str(TOOLS / "pack_gguf.py"), "--model", str(setdir), "--out", str(pack)]+ if ENGRAM_DIR.exists():+ cmd += ["--engram", str(ENGRAM_DIR.parent)]+ if DENSE_WOA.exists():+ cmd += ["--dense-override", str(DENSE_WOA)]+ if DENSE_HEAD.exists():+ cmd += ["--dense-override", str(DENSE_HEAD)]+ cmd += ["--skip-tensors", r"blk\.\d+\.attn_output_a\.weight"]+ if DSPARK.exists():+ cmd += ["--dspark-gguf", str(DSPARK), "--dspark-noise-token", "128799", "--dspark-markov-rank", "256"]+ log(" pack: %s" % " ".join(cmd[2:]))+ r = run(cmd, cwd=str(PORT))+ (PORT / "runs").mkdir(exist_ok=True)+ (PORT / "runs" / ("pack_%s_%s.log" % (set_name, commit[:7]))).write_text((r.stdout or "") + (r.stderr or ""))+ if r.returncode != 0:+ report.append("- pack FAILED (rc %d), see runs/pack_%s_%s.log" % (r.returncode, set_name, commit[:7]))+ return None+ report.append("- pack: %s" % pack)+ return pack+++def tensor_names(setdir: pathlib.Path) -> list[str]:+ sys.path.insert(0, str(TOOLS))+ try:+ import pack_gguf+ except Exception:+ return []+ names = []+ for s in sorted(setdir.glob("*.gguf")):+ try:+ g = pack_gguf.Gguf(s)+ names += [t[0] for t in g.tensors]+ g.close()+ except Exception:+ continue+ return names+++def pack_info(pack: pathlib.Path) -> str:+ r = run(["bash", "-lc", ". /data/dsv41-port/env.sh; cd /data/dsv41-port && ./build/strata-dsv41 --pack-info --pack %s"+ % pack])+ return (r.stdout or "") + (r.stderr or "")+++def eval_pack(pack: pathlib.Path, tag: str) -> list[str]:+ r = run(["bash", "-lc", "cd /data/dsv41-port && bash session16.sh %s" % pack])+ (PORT / "runs" / ("eval_%s.log" % tag)).write_text((r.stdout or "") + (r.stderr or ""))+ rows = [l for l in (r.stdout or "").splitlines() if l.startswith("s16_") or l.startswith("rc=")]+ return rows+++def speed_smoke(pack: pathlib.Path, tag: str) -> list[str]:+ out = []+ ids = (PORT / "prompts/sa_code.ids").read_text()[:300]+ for prompt, name in ((ids, "code"),):+ for dspark in ("off", "on"):+ cmd = ("cd /data/dsv41-port && . ./env.sh && ./build/strata-dsv41 --pack %s --tp 1 --threads off --gpus 0 "+ "--max-context 2048 --vision off --dspark %s --expert-profile none --ids '%s' --n 32"+ % (pack, dspark, prompt))+ t0 = time.time()+ r = run(["bash", "-lc", cmd])+ dt = time.time() - t0+ m = re.search(r"decode:.*?([\d.]+) tok/s", (r.stderr or ""))+ out.append("| %s | %s | %s | %s | %.1f |" % (name, dspark, "ok" if r.returncode == 0 else "FAILED",+ m.group(1) if m else "-", dt))+ return out+++def main() -> int:+ ap = argparse.ArgumentParser()+ ap.add_argument("--commit", required=True)+ ap.add_argument("--sets", default="2.00bpw")+ ap.add_argument("--root", default=str(ROOT))+ ap.add_argument("--no-download", action="store_true")+ ap.add_argument("--no-pack", action="store_true")+ ap.add_argument("--no-eval", action="store_true")+ ap.add_argument("--no-speed", action="store_true")+ a = ap.parse_args()+ commit = a.commit+ root = pathlib.Path(a.root)+ rep_dir = REPORTS / commit[:7]+ rep_dir.mkdir(parents=True, exist_ok=True)+ report: list[str] = ["# IST-DAS intake, commit %s" % commit, "", "run %s" % time.strftime("%Y-%m-%dT%H:%M:%S%z"), ""]++ log("listing %s at %s" % (REPO, commit[:7]))+ tree = repo_tree(commit)+ (rep_dir / "tree.txt").write_text("\n".join("%s %d" % (p, s) for p, s in tree))+ report += ["## Files at this commit", "", "| path | GB |", "|---|---|"]+ report += ["| %s | %.2f |" % (p, s / 1e9) for p, s in tree]+ report.append("")++ rc = 0+ for set_name in [s.strip() for s in a.sets.split(",") if s.strip()]:+ report += ["## Set %s" % set_name, ""]+ files = [(p, s) for p, s in tree if p.startswith(set_name + "/") and p.endswith(".gguf")]+ # 1. download what is missing or the wrong size+ if not a.no_download:+ for p, size in files:+ dest = local_path(root, p)+ if dest.exists() and dest.stat().st_size == size:+ report.append("- present: %s (%.2f GB)" % (p, size / 1e9))+ continue+ log("downloading %s (%.2f GB)" % (p, size / 1e9))+ d = download(commit, p, root)+ if d != 0:+ report.append("- DOWNLOAD FAILED: %s" % p)+ rc = 1+ else:+ report.append("- downloaded: %s (%.2f GB)" % (p, size / 1e9))+ # 2. hash + dedupe+ sums = root / ("SHA256SUMS-%s-%s.txt" % (commit[:7], set_name))+ lines = []+ for p, size in files:+ dest = local_path(root, p)+ if not dest.exists():+ continue+ lines.append("%s %s" % (sha256_file(dest), p))+ sums.write_text("\n".join(lines) + "\n")+ report.append("- sha256: %s" % sums)+ # 3. headers+ shards = [local_path(root, p) for p, _ in files if local_path(root, p).exists()]+ if shards:+ try:+ types = tensor_types(shards)+ report += ["", "| type | tensors | GB | device kernel |", "|---|---|---|---|"]+ for t, (c, b) in sorted(types.items()):+ report.append("| %s (%d) | %d | %.2f | %s |" % (+ TYPE_NAMES.get(t, "?"), t, c, b / 1e9, "yes" if t in DEVICE_TYPES else "**NO - CPU lane**"))+ report.append("")+ except Exception as e: # noqa: BLE001+ report.append("- header scan failed: %r" % (e,))+ # 4. pack+ if not a.no_pack:+ pack = pack_set(commit, set_name, root, report)+ else:+ pack = None+ # 5. pack-info + eval + speed+ if pack is not None:+ pi = pack_info(pack)+ report += ["### --pack-info", "", "```", pi.strip()[-3000:], "```", ""]+ if "MISSING" in pi:+ report.append("- **pack-info reports MISSING files: this pack cannot load**")+ rc = 1+ if not a.no_eval:+ log("eval %s" % pack)+ rows = eval_pack(pack, "%s_%s" % (set_name, commit[:7]))+ report += ["### quality (4995 positions, code/agent/english)", "", "```", "\n".join(rows) or "(no output)", "```", ""]+ if not a.no_speed:+ log("speed smoke %s" % pack)+ rows = speed_smoke(pack, "%s_%s" % (set_name, commit[:7]))+ report += ["### speed smoke (32 tokens, one card)", "", "| prompt | dspark | rc | decode tok/s | wall s |",+ "|---|---|---|---|---|"] + rows + [""]+ report.append("")++ out = rep_dir / ("%s.md" % ",".join(s.strip() for s in a.sets.split(",")))+ out.write_text("\n".join(report) + "\n")+ log("report: %s" % out)+ print("REPORT %s" % out)+ return rc+++if __name__ == "__main__":+ sys.exit(main())