Upload app_deepseek.py

app_deepseek.py

@@ -388,11 +388,17 @@ SEARCH_LANGUAGE_CYCLE = ["arabic", "english"]
 # "llamacpp" (default): llama-server + GGUF, on GPU (CUDA build) or CPU. llama.cpp cannot use TPU.
 # "hf": the old transformers path (only kept for reference; on TPU it produced ~3 chars/s and hung).
 BACKEND = os.environ.get("GIJL_BACKEND", "llamacpp").strip().lower()
-GGUF_REPO = os.environ.get("GIJL_GGUF_REPO", "unsloth/DeepSeek-V4-Flash-GGUF")
-GGUF_QUANT = os.environ.get("GIJL_GGUF_QUANT", "UD-Q2_K_XL")   # highest quant unsloth ships; choose_quant() steps this down
-                                                           # automatically to whatever fits the detected VRAM/RAM
-                                                           # budget (see _memory_budget_gb) -- set GIJL_GGUF_QUANT
-                                                           # explicitly to pin a smaller quant instead.
+GGUF_REPO = os.environ.get("GIJL_GGUF_REPO", "unsloth/DeepSeek-V4-Flash-0731-GGUF")
+# v6 bugfix: this repo ships no plain "Q8_0" quant directory -- only "UD-"-prefixed dynamic quants
+# (UD-IQ1_M .. UD-Q8_K_XL) plus a top-level "dspark-DeepSeek-V4-Flash-0731-Q8_0.gguf" speculative-decoding
+# DRAFT model whose filename happens to end in "-Q8_0". Before the _group_gguf_files() exclusion fix below,
+# that draft file was mis-grouped as if it were the real "Q8_0" main-model quant and downloaded/loaded via
+# -m, which llama-server then refused to initialize standalone ("dflash requires ctx_other to be set").
+# Pinned to "UD-IQ3_XXS" (~104.2GB across its 4 shards, per the repo's own file listing) by request --
+# the closest real quant this repo ships to a ~103-104GB file size. choose_quant() still steps it down
+# automatically if it somehow doesn't fit the detected VRAM/RAM budget (see _memory_budget_gb); set
+# GIJL_GGUF_QUANT to override explicitly.
+GGUF_QUANT = os.environ.get("GIJL_GGUF_QUANT", "UD-IQ3_XXS")
 LLAMA_CTX = int(os.environ.get("GIJL_CTX", "12288"))            # tokens; KV cache is ~2-4GB at this size
 LLAMA_LOAD_TIMEOUT_S = 1800
 LLAMA_BIN_DIR = os.path.join(WORK_DIR, "bin")                   # CUDA build cache (synced to the hub, see CACHE_LLAMA_BINARY)
@@ -1255,7 +1261,15 @@ def _group_gguf_files(siblings) -> dict:
     for name, size in siblings:
         base = os.path.basename(name)
         low = base.lower()
-        if not low.endswith(".gguf") or any(x in low for x in ("mmproj", "dflash", "assistant", "draft")):
+        if not low.endswith(".gguf") or any(x in low for x in ("mmproj", "dflash", "dspark", "assistant", "draft")):
+            # v6 bugfix: "dspark" added -- Unsloth ships DSpark speculative-decoding DRAFT models
+            # (e.g. "dspark-DeepSeek-V4-Flash-0731-Q8_0.gguf") that are NOT standalone models; llama-server
+            # refuses to load one via -m alone ("dflash requires ctx_other to be set") because it expects to be
+            # paired with a real main-model quant via --model-draft. Without this exclusion, the quant-tag
+            # regex below still matches e.g. "-Q8_0" at the end of the draft file's name and silently groups
+            # it as if it were a legitimate "Q8_0" main-model quant, so it gets selected and downloaded instead
+            # of the real model whenever a repo (like unsloth/DeepSeek-V4-Flash-0731-GGUF) has no quant
+            # directory actually named "Q8_0" -- causing a hard crash at startup.
             continue
         m = _QUANT_RE.search(_SHARD_RE.sub("", base[:-5]))
         if not m: