Upload app_deepseek.py

app_deepseek.py

@@ -293,6 +293,23 @@ changes where the tokenizer/template are read from). Kept/strengthened as the re
 its 1M context lets a scenario be grounded in a FULL paper or advisory instead of a MAX_PROMPT_CHARS-
 truncated excerpt. See WHAT CHANGED IN v8 for the real enable_thinking=False switch and native DSML
 tool-call fallback this chat_template.jinja also made possible.
+v8.1 (field report, CPU-only): llama-server itself (not this file) crashed on startup --
+GGML_ASSERT(*cur_backend_id != -1) failed in ggml_backend_sched_split_graph, during context/graph
+reserve, before any chat() call -- with -ngl 0 on this exact GGUF. DeepSeek-V4 support in llama.cpp was
+still WIP/fork-based as of its own GitHub support thread (multiple different per-backend crashes already
+reported there, none of them this exact one) -- the signature (no backend can run some op in the graph)
+matches "this architecture's CPU backend coverage is still incomplete", not a bug in this script,
+confirmed by the earlier pipeline stages (model/tokenizer/chat_template.jinja all downloaded and loaded
+fine; the crash is inside the llama-server binary itself). Two things to try, cheapest first: (1) a
+PLAIN quant instead of an Unsloth Dynamic one -- GIJL_GGUF_QUANT=Q2_K (85.7GB) or Q3_K_S (101GB) --
+since UD- quants mix per-layer precision/tensor types a brand-new architecture's CPU graph-splitter may
+not handle as uniformly; (2) GIJL_LLAMA_CPU_TAG=<tag> (new in v8.1, see download_llama_cpu_release()) to
+pin a specific llama.cpp release instead of always "latest", to confirm whether this is build-specific.
+If neither works, this is upstream, not local: either wait for llama.cpp's CPU backend for this
+architecture to mature (GIJL_LLAMA_SERVER is unset by default, so a later run automatically tries
+whatever "latest" is by then -- no file change needed), or point GIJL_LLAMA_SERVER (already supported,
+see ensure_llama_server()) at a self-built binary from an experimental third-party CPU-capable fork --
+unofficial, so verify it yourself before trusting it with real data.
 FOUR sibling files (glm/minimax/deepseek/kimi/qwen minus this one) run the same pipeline, each specialized
 differently -- see WHAT CHANGED IN v7 and WHAT CHANGED IN v8 at the top of this file for the full table and the
 evidence behind each choice. all five write into the SAME dataset concurrently -- see MODEL_TAG and the v6 note above for how that stays race-free. The chat template + tokenizer come from the
@@ -1553,15 +1570,27 @@ def download_llama_cpu_release():
     arch = "arm64" if platform.machine().lower() in ("aarch64", "arm64") else "x64"
     base = "https://github.com/ggml-org/llama.cpp/releases"
     hdr = {"User-Agent": _CRAWLER_UA}
-    tags = []
-    try:
-        latest = requests.get(f"{base}/latest", timeout=30, headers=hdr).url.rstrip("/").rsplit("/", 1)[-1]
-        nt = requests.get(f"{base}/download/{latest}/nightly-tag.txt", timeout=30, headers=hdr)
-        if nt.status_code == 200 and nt.text.strip():
-            tags.append(nt.text.strip().split()[0])
-        tags.append(latest)
-    except requests.RequestException as e:
-        raise RuntimeError(f"cannot reach GitHub releases: {_redact(str(e))[:200]}")
+    pinned_tag = os.environ.get("GIJL_LLAMA_CPU_TAG", "").strip()
+    if pinned_tag:
+        # v8: bisecting knob, opt-in only (default behavior below is unchanged) -- added because DeepSeek-V4-Flash
+        # crashed llama-server on CPU with GGML_ASSERT(*cur_backend_id != -1) in ggml_backend_sched_split_graph
+        # during context/graph setup, a signature consistent with "this architecture's CPU backend coverage in
+        # llama.cpp is still incomplete" rather than anything wrong in this file -- DeepSeek-V4 support was still
+        # WIP/fork-based in llama.cpp as of its own GitHub support thread. "latest" always means whatever today's
+        # CI build is, so there is no way to test a different one without this. EXCLUSIVE on purpose (no silent
+        # fallback to latest): the point is to know for certain which build actually ran.
+        print(f"[llama] GIJL_LLAMA_CPU_TAG={pinned_tag} set -- using this exact release instead of latest", flush=True)
+        tags = [pinned_tag]
+    else:
+        tags = []
+        try:
+            latest = requests.get(f"{base}/latest", timeout=30, headers=hdr).url.rstrip("/").rsplit("/", 1)[-1]
+            nt = requests.get(f"{base}/download/{latest}/nightly-tag.txt", timeout=30, headers=hdr)
+            if nt.status_code == 200 and nt.text.strip():
+                tags.append(nt.text.strip().split()[0])
+            tags.append(latest)
+        except requests.RequestException as e:
+            raise RuntimeError(f"cannot reach GitHub releases: {_redact(str(e))[:200]}")
     last_err = "no candidate tag"
     for tag in tags:
         dest = os.path.join(LLAMA_CPU_DIR, tag)