frank-default-large: cap CTX_SIZE at 32768 (author quick-start) — 87 GiB fully-resident model + context buffer OOMs HIP heap (50.8 GiB at 131072, 103 GiB at 262144)

AGENTS.md

@@ -27,7 +27,7 @@ host.
 | Target | Runtime | Service | Model | Port |
 |--------|---------|---------|-------|------|
 | Dankserver | llama.cpp (upstream b10121) | frank-default-small | HauhauCS Qwen3.6-35B-A3B-Uncensored-Aggressive Q6_K_P + ngram-mod, 36 MoE experts on CPU | 7770 |
-| Dankclaw | charlie12345/ROCmFPX llama.cpp (build-qwen38) | frank-default-large | agentionai Qwen3.8-Flash-Next-ROCmFP4-FAST-imatrix (87 GiB Q4_0_ROCMFP4_FAST, PLE joined table in v2 GGUF), ngram-mod spec (MTP head on disk, incompatible with this build), 131K context / 1 slot, reasoning on | 7770 |
+| Dankclaw | charlie12345/ROCmFPX llama.cpp (build-qwen38) | frank-default-large | agentionai Qwen3.8-Flash-Next-ROCmFP4-FAST-imatrix (87 GiB Q4_0_ROCMFP4_FAST, PLE joined table in v2 GGUF), ngram-mod spec (MTP head on disk, incompatible with this build), 32K context / 1 slot, reasoning on | 7770 |
 
 History: the 2026-08-17 redo retired the pipeline-parallel split and the
 FastFlowLM NPU endpoint. The 2026-08-18 swap retired the vLLM service on
@@ -293,12 +293,13 @@ VRAM-resident, no sidecar, no `--ngram-on-disk`. The model's own MTP draft head
 quant matched to target) is on disk but does **not** load on this binary
 (loader requires a full per-layer `blk.<il>.hc_attn_norm.weight` set; the head
 carries only `blk.48`) — speculative decoding runs ngram-mod only. Vision
-tower (f16) ships as mmproj/. It runs 131072 context / 1
-slot (capped: the GPU-resident 87 GiB model + 262144 f16 KV + compute
-buffer OOMs the SVM pool — verified live, cudaMalloc 103176 MiB), b 4096 / ub 4096, t 16 / tb 16, f16 KV cache, `--fit off` (auto-fit
-segfaults on qwen4exp hyper-connection config), `--flash-attn off` (required:
-fa on aborts graph build on the QSA indexer path — F32 `ggml_fill` on the F16
-kq mask, ggml.c:5371), and reasoning on (Sharp
+tower (f16) ships as mmproj/. It runs 32768 context / 1
+slot (author quick-start; the fully GPU-resident 87 GiB model + f16 KV +
+compute buffer OOMs the SVM pool beyond that — verified live: cudaMalloc
+fails at 131072/262144), b 4096 / ub 4096, t 16 / tb 16, f16 KV cache,
+`--fit off` (auto-fit segfaults on qwen4exp hyper-connection config),
+`--flash-attn off` (required: fa on aborts graph build on the QSA indexer
+path — F32 `ggml_fill` on the F16 kq mask, ggml.c:5371), and reasoning on (Sharp
 template, budget -1, `--reasoning-format deepseek`; chat kwargs keep
 `preserve_thinking` + `preserve_reasoning`). Env
 `scripts/frank-default-large.env`, serve `scripts/serve-frank-large.sh`.

scripts/frank-default-large.env

@@ -60,11 +60,14 @@ LISTEN_PORT=7770
 # hyper-connection/indexer config; -ngl 999 puts every layer on the
 # SVM-backed GPU pool.
 #
-# CTX_SIZE is capped at 131072: the fully GPU-resident 87 GiB model plus a
-# 262144-token f16 KV cache and the worst-case compute buffer exceed the
-# Strix Halo SVM pool — verified live (cudaMalloc 103176 MiB OOM at
-# 262144). 131072 is the proven STRIX-family value on this box.
-CTX_SIZE=131072
+# CTX_SIZE is capped at 32768 (the author's own quick-start). This model is
+# fully GPU-resident (87.06 GiB) — the PLE n-gram table no longer streams
+# from RAM like STRIX_LEAN did, so there is no room for a huge context
+# buffer: HIP heap is 119.3 GiB, and the context buffer alone is 50.8 GiB
+# at 131072 and 103 GiB at 262144 (verified live:
+# `cudaMalloc … failed: out of memory` on first graph alloc). At 32768 the
+# context buffer drops to ~13 GiB (~100 GiB total) and loads clean.
+CTX_SIZE=32768
 PARALLEL=1
 BATCH_SIZE=4096
 UBATCH_SIZE=4096

scripts/serve-frank-large.sh

@@ -45,7 +45,7 @@ fi
 : "${ALIAS:=frank-default-large}"
 : "${BIND_ADDR:=0.0.0.0}"
 : "${LISTEN_PORT:=7770}"
-: "${CTX_SIZE:=131072}"
+: "${CTX_SIZE:=32768}"
 : "${PARALLEL:=1}"
 : "${BATCH_SIZE:=4096}"
 : "${UBATCH_SIZE:=4096}"

serverAppConfigs-docs/LLM/agentionai-fast-imatrix.md

@@ -89,7 +89,7 @@ llama-server \
   --host 0.0.0.0 --port 7770 --jinja \
   --device ROCm0 --gpu-layers 999 \
   --split-mode none --fit off --flash-attn off \
-  --ctx-size 131072 --parallel 1 \
+  --ctx-size 32768 --parallel 1 \
   --batch-size 4096 --ubatch-size 4096 \
   --cache-type-k f16 --cache-type-v f16 \
   --threads 16 --threads-batch 16 \
@@ -119,12 +119,13 @@ Key choices and deviations from the CIRU deployment:
   mask and aborts graph build (`GGML_ASSERT(a->type == GGML_TYPE_F32)`,
   ggml.c:5371, backtrace through `build_attn_qsa`). Verified live on this
   v2 main — loads and serves fine with fa off.
-- Context capped at 131072 × 1 slot. The 262144 window does NOT fit: the
-  fully GPU-resident 87 GiB model plus a 262144-token f16 KV cache plus the
-  worst-case compute buffer exceeds the Strix Halo SVM pool (verified live:
-  `cudaMalloc 103176 MiB failed: out of memory` on first graph alloc at
-  262144). 131072 is the proven STRIX-family value on this box. (The author's
-  quick-start is 32768 — the full 131K is comfortable.)
+- Context capped at **32768 × 1 slot — the author's own quick-start**. This
+  model is fully GPU-resident (87.06 GiB): the PLE n-gram table sits inside
+  the GGUF as one tensor, unlike STRIX_LEAN which streamed it from RAM.
+  HIP heap is 119.3 GiB; the context buffer is 50.8 GiB at 131072 and
+  103 GiB at 262144, so either window OOMs (`cudaMalloc … failed: out of
+  memory`, verified live on first graph alloc). At 32768 the buffer is
+  ~13 GiB (~100 GiB total) and loads clean.
 - MTP is disabled (ngram-mod only) — see the MTP section below.
 
 ## MTP draft head — INCOMPATIBLE on build-qwen38