frank-medium: serve the new R9700s, retire frank-small

The RTX 3090 left Dankserver; two AMD Radeon AI PRO R9700 cards (Navi 48,
gfx1201, 32,624 MiB each) replaced it. FreeToken/CUDA cannot run there at all,
so the `frank-small` service is retired and `frank-medium` takes the endpoint.

New service: charlie12345/ROCmFPX llama.cpp (built for gfx1201 against the
fork's self-contained ROCm 7.14 toolchain, so the host's stale /opt/rocm-6.3.3
never participates) serving kingjones777/Qwen3.8-27B-ROCmFPX-Q8_0-AGENT-GGUF --
a Q8_0_ROCMFPX_AGENT (ftype 115) checkpoint that stock llama.cpp cannot load.
Placeholder config: --split-mode tensor across both cards, 131072 context,
1 slot, f16 KV, MTP draft head on ROCm0.

deploy-remote.sh grows retire_root_llm_small(): the CUDA-era unit is stopped,
disabled and moved to ~/.AGENTTRASH/20260915-frank-small-retire/host-units
before frank-medium starts, because `systemctl disable` has been observed to
delete the unit file itself on this host (2026-08-26 incident).

The retired frank-small repo files (env, unit, serve script) move to
scripts/.AGENTTRASH/, and the planner, dispatcher, service allow-list and the
deployment test seams now name frank-medium.

Follow-up (same session, measured): llama-bench on the host with the correct
multi-device syntax (-dev ROCm0/ROCm1) overturned the tensor choice:

  single R9700   pp4096 1160.9 / tg128 19.42  (d32768: 512 / 18.4)
  layer  2x      pp4096 1991.2 / tg128 18.76  (d32768: 859 / 17.6)
  tensor 2x       pp4096  271.8 / tg128 26.26 (d32768: 229 / 25.1)

This build has no RCCL, so tensor mode falls back to a "meta-backend
butterfly" allreduce and prefill collapses 7x. Served path with MTP:
layer 35.7-46.2 tok/s vs tensor 28.7. SPLIT_MODE=layer, draft pinned to
ROCm1 (the card holding the last target layers). Served prefix cache with
the draft head verified live: turn 2 reprocessed 22 tokens instead of 949.

scripts/frank-medium.env

@@ -25,12 +25,20 @@
 # MTP draft acceptance (0.953 vs 0.911) and that is the whole point of this
 # checkpoint.  Model card documents 131,072 context.
 #
-# Parallelism (--split-mode tensor): tensor parallel splits the weights AND
+# Parallelism (--split-mode layer, operator call 2026-09-15 "pipeline
-# the KV cache across both cards through the fused "meta device".  It is the
+# parallelism"): layers are distributed across the two cards in order and
-# only split mode here that gives each card half the KV cache; layer split
+# activations hop PCIe at the boundary.  Measured with llama-bench
-# would give each card the KV of the layers it owns, with one PCIe handoff
+# (pp4096 / tg128 / same at depth 32768, f16 KV, fa on, 3 reps):
-# per token per boundary.  TENSOR mode's own constraints, all met below:
+#
-# flash attention ON, non-quantized KV cache, and no auto-fit.
+#   single R9700   1160.9 / 19.42   (fits: 26.27 GiB weights, ~5 GiB left)
+#   layer  2x      1991.2 / 18.76   859.1 / 17.64
+#   tensor 2x       271.8 / 26.26   228.6 / 25.09
+#
+# Tensor parallel wins decode (+40 %) but collapses prefill 7x: this build
+# has no RCCL (llama.cpp logs "AllReduce init failed; falling back to
+# meta-backend butterfly"), so every layer pays a slow cross-card reduction.
+# Agentic traffic is prefill-heavy, so layer split wins the workload.  Revisit
+# if the fork ships a working RCCL tensor path.
 #
 # WARNING: do not name any variable TEMP, HOST, or GPU_MAX_HEAP_SIZE. The
 # ROCm HIP runtime reads those bare env vars (see the 2026-08-13 TEMP
@@ -59,20 +67,20 @@ LISTEN_PORT=7770
 # can drift onto shared system RAM; DEVICES is what llama.cpp splits across.
 HIP_VISIBLE_DEVICES=0,1
 DEVICES=ROCm0,ROCm1
-SPLIT_MODE=tensor
+SPLIT_MODE=layer
 
-# TENSOR_SPLIT empty = an even split between the two cards.
+# TENSOR_SPLIT empty = layers distributed by free VRAM (equal here).
 TENSOR_SPLIT=
 
 NGL=999
-# Tensor mode ignores auto-fit, and auto-fit has no business resizing a
+# Auto-fit has no business resizing a context we set explicitly.
-# context we set explicitly.
 FIT=off
 FLASH_ATTN=on
 
-# 131,072 tokens, one slot (the model card's documented window).  Tensor mode
+# 131,072 tokens, one slot (the model card's documented window; the GGUF
-# puts half the KV cache on each card, so the window is not limited by a
+# reports n_ctx_train 262144, untested by the model author).  Layer split
-# single card's free memory.
+# keeps each layer's KV on the card that owns the layer, so the window is
+# bounded by per-card free memory: ~5.6 GiB of KV per card at 131072 f16.
 CTX_SIZE=131072
 PARALLEL=1
 
@@ -83,8 +91,7 @@ UBATCH_SIZE=512
 NUM_THREADS=8
 THREAD_BATCH=8
 
-# f16 K/V: tensor mode refuses a quantized KV cache, and an agent checkpoint
+# f16 K/V: an agent checkpoint is the wrong place to spend KV precision.
-# is the wrong place to spend KV precision anyway.
 CACHE_TYPE_K=f16
 CACHE_TYPE_V=f16
 CTX_CHECKPOINTS=8
@@ -122,11 +129,10 @@ SPEC_DRAFT_TYPE_K=f16
 SPEC_DRAFT_TYPE_V=f16
 SPEC_DRAFT_POLL=1
 SPEC_DRAFT_POLL_BATCH=1
-# The draft head must run on ONE device.  With --split-mode tensor the draft
+# The draft head rides the card that holds the target model's LAST layers:
-# otherwise joins the meta-device allocation and the KV-cache build aborts:
+# the MTP head consumes the final hidden state, so pinning it to ROCm1 keeps
-#   ggml/src/ggml-backend.cpp:119: GGML_ASSERT(buffer) failed
+# the per-token hop off the wire.  (Under the abandoned tensor split the
-#   <- ggml_backend_meta_alloc_ctx_tensors_from_buft <- llama_kv_cache ctor
+# draft must also be pinned to exactly one device or the meta-device KV
-# (measured 2026-09-15 on the built runtime: identical flags are healthy with
+# allocation aborts: GGML_ASSERT(buffer) in ggml-backend.cpp, measured
-# --spec-draft-device ROCm0 and abort with it empty).  ROCm0 is the card that
+# 2026-09-15.)
-# also holds the target model's main GPU.
+SPEC_DRAFT_DEVICE=ROCm1
-SPEC_DRAFT_DEVICE=ROCm0

scripts/serve-frank-medium.sh

@@ -27,11 +27,13 @@
 # what makes the MTP draft land ~0.95 of its guesses.  A stock llama.cpp
 # cannot load this GGUF at all - Q8_0_ROCMFPX exists only in this fork.
 #
-# Parallelism: --split-mode tensor across both cards.  Tensor parallel
+# Parallelism: --split-mode layer across both cards (operator call
-# splits the weights AND the KV cache across the participating devices via
+# 2026-09-15).  Layers are distributed in order and activations hop PCIe at
-# the fused "meta device" (unlike -sm layer, which pipelines layers and
+# the boundary; measured llama-bench: layer 1991/18.8 tok/s (pp4096/tg128)
-# keeps each layer's KV on its own card).  It requires flash attention and
+# vs tensor 272/26.3 and single-card 1161/19.4.  Tensor wins decode +40 %
-# a non-quantized KV cache, and it is disabled from auto-fit (--fit off).
+# but its prefill collapses 7x on this build (no RCCL: "AllReduce init
+# failed; falling back to meta-backend butterfly"), so layer split wins the
+# prefill-heavy agentic workload.
 #
 # Tunables come from frank-medium.env unless FRANK_MEDIUM_ENV_FILE
 # overrides that path.