@miller-tech/uap 1.215.2 → 1.217.0

package/config/llama-profiles/gemma4-26b-a4b-mtp.env

@@ -124,7 +124,7 @@ LLAMA_LOG_FILE=/home/cogtek/llama.cpp/llama-server.log
 LLAMA_N_PREDICT=16384
 
 # Gemma 4 speaks its own <|turn>/<|tool_call> DSL, NOT ChatML. Use the
-# model's embedded template — the qwen3.5-enhanced.jinja used by the Qwen
+# model's embedded template — the qwen-sharp.jinja used by the Qwen
 # profile emits a format Gemma was never trained on and silently destroys
 # tool calling.
 LLAMA_CHAT_TEMPLATE_FILE=embedded

package/config/llama-profiles/qwen36-35b-a3b.env

@@ -57,9 +57,58 @@ LLAMA_SLOT_SAVE_PATH=/home/cogtek/.cache/inference/kv-slots
 LLAMA_LOG_FILE=/home/cogtek/llama.cpp/llama-server.log
 LLAMA_N_PREDICT=16384
 
-# Qwen3.6 is ChatML with the qwen3_coder tool-call format; the proxy's parsers
-# are built around exactly this template. Do NOT switch it to "embedded".
-LLAMA_CHAT_TEMPLATE_FILE=/home/cogtek/dev/miller-tech/universal-agent-protocol/tools/agents/config/qwen3.5-enhanced.jinja
+# CHAT TEMPLATE (2026-08-20): switched from qwen3.5-enhanced.jinja to
+# qwen-sharp.jinja -- peculiar-ragdoll/Qwen-Sharp-Chat-Templates
+# (template_version qwen3.8-froggeric-v22.1), vendored verbatim at
+# tools/agents/config/qwen-sharp.jinja. It targets Qwen 3.5/3.6/3.8 and is the
+# template for ALL qwen models here. Pinned artifact -- both are asserted in
+# test/chat-template-system-message.test.ts so a silent re-vendor fails CI:
+#   upstream revision  3dc34df52c63dd22ada21f96435e069deaa8d7da (2026-08-18)
+#   file sha256        d1f22a89eac3609dcfaa7b471b1f7d23bee2f084d275d26f4f8231d1d7908f4e
+#
+# WHAT IT FIXES: qwen3.5-enhanced.jinja hard-raised
+# 'No user query found in messages.' -- a server-side failure for the whole
+# request -- whenever no user turn survived its backwards scan, i.e. when every
+# user turn was a <tool_response> or when the transcript was system+assistant
+# only. Both shapes occur in normal agent loops. qwen-sharp renders them.
+#
+# WHAT STAYS COMPATIBLE (rendered and compared, not assumed):
+#   * the tool-call contract the model is told to EMIT is unchanged:
+#     <tool_call><function=NAME><parameter=K>V</parameter></function></tool_call>.
+#     anthropic_proxy.py's _HERMES_FUNCTION_RE / _TOOL_CALL_XML_RE parse it
+#     unchanged. The default is selected at render time by `_tool_format`, so
+#     the test asserts on RENDERED output -- the file contains a JSON emitter
+#     too, and a grep cannot tell which one is live.
+#   * the generation prompt matches the old template in BOTH enable_thinking
+#     states, so --reasoning auto --reasoning-format deepseek is unaffected.
+#   * a mid-conversation system message renders inline in both templates.
+#
+# WHAT ACTUALLY DIFFERS -- known and accepted:
+#   * HISTORY rendering of prior tool calls differs in whitespace only: with
+#     string-form `arguments` (the proxy's path -- it json.dumps them) the old
+#     template emitted "}\n</function>", the new one "}</function>"; and the
+#     separator between two tool calls in one assistant turn went "\n" ->
+#     "\n\n". Neither changes what the parsers accept.
+#   * the injected tool-calling instruction block was rewritten upstream: it now
+#     requires a <think> block before each tool call and forbids prose before
+#     the call. Behavioural, deliberate, not a wire change.
+#   * a system message containing an IMAGE now raises. Unreachable via the
+#     proxy, which flattens system to a text string before dispatch
+#     (anthropic_proxy.py ~4709). Direct-to-:8080 OpenAI clients can hit it.
+#   * unknown message roles no longer raise; they render as "[role]: content".
+#   * prior <think> blocks are retained across turns. Upstream sells this as a
+#     prompt-cache win; on the PROXY path it is a no-op, because the proxy
+#     already strips <think> from prior assistant turns on purpose
+#     (anthropic_proxy.py ~4725: "Stripping breaks the copy cycle"). On the
+#     direct path it re-introduces exactly that cycle -- pass
+#     chat_template_kwargs.preserve_thinking=false there if it shows up.
+#
+# chat_template_kwargs.reasoning_effort accepts
+# none|minimal|low|medium|high|xhigh|max. NOTE: 'none' forces thinking OFF,
+# overriding enable_thinking=true.
+# Do NOT switch this to "embedded" -- the model's own baked-in template still
+# carries the 'System message must be at the beginning.' raise.
+LLAMA_CHAT_TEMPLATE_FILE=/home/cogtek/dev/miller-tech/universal-agent-protocol/tools/agents/config/qwen-sharp.jinja
 
 # --cache-ram sizing — measured, not guessed
 # The host-RAM prompt cache holds KV prefixes so a follow-up turn re-uses its

package/config/llama-profiles/qwen38-27b-mtp.env

@@ -0,0 +1,155 @@
+# =====================================================================
+# Qwen 3.8 27B (dense, UD-IQ4_XS) @ 131072 ctx/slot / 2 slots, native MTP spec
+# decoding on llama.cpp master (f466cfa38, 2026-08-20) — 2026-08-20
+# Service: uap-llama-server.service -> scripts/run-llama-server-continuity.sh
+# Install: cp to ~/.config/uap/llama-server.env
+#
+# WHY THIS BUILD: previous profile (qwen36-35b-a3b.env, kept alongside this
+# file) ran a stale mtp-port worktree binary ~5 months behind upstream, with
+# spec decoding off. Investigated wiring in the operator's local "dflash"
+# tree-kernel / DDTree branches for tree-mode speculative verification on
+# hybrid SSM+attention models; found upstream (commit 4e732e0a6, "llama:
+# allow partial seq_rm for GDN models for speculative decoding") already
+# merged the same core mechanism natively, and the remaining DDTree/tree-
+# topology piece has no upstream equivalent and would need fresh kernel R&D,
+# not a port -- deferred rather than shipped half-verified. This profile is
+# instead plain upstream master + native `draft-mtp`, which measured 60-91%
+# draft acceptance and no corruption/crashes across single, 5x-consecutive,
+# multi-turn, and -np 2 concurrent validation queries.
+#
+# WHY MTP WORKS HERE (unlike the old profile's note below): Qwen3.8-27B-
+# UD-IQ4_XS.gguf is a DENSE qwen35-arch model with a native MTP head baked
+# into the checkpoint (blk.64.nextn.* tensors) -- confirmed via `gguf`
+# tensor dump. The old profile's "MTP is a no-op" finding was specifically
+# about a MoE (A3B) model; it does not apply here. No -md draft model file
+# is used or needed -- draft-mtp reads the head from LLAMA_MODEL itself.
+#
+# CONTEXT BUDGET: --ctx-size is the TOTAL across slots, /props reports the
+# PER-SLOT figure. 262144 / 2 slots = 131072 tokens/slot (2026-08-20: raised
+# from 65536/slot per operator request). See
+# docs/guides/QWEN36_LLAMACPP.md#measured-qwen38-27b-dense--native-draft-mtp-rtx-3090-24-gb
+# for measured VRAM headroom at both tiers on this card (24 GiB) -- at
+# 131072/slot only ~0.95 GiB is free, verified stable but thin; drop back to
+# 65536/slot first if instability appears under real concurrent load. Keep
+# PROXY_CONCURRENCY_LIMIT == slot count and PROXY_CONTEXT_WINDOW == per-slot
+# context (anthropic-proxy.env; re-derive from /props if this changes).
+# =====================================================================
+LLAMA_BIN=/home/cogtek/llama.cpp/.worktrees/dflash-mtp-upgrade/build-cuda/bin/llama-server
+LLAMA_MODEL=/home/cogtek/Downloads/Qwen3.8-27B-UD-IQ4_XS.gguf
+# CONFIRMED INCOMPATIBLE 2026-08-20: mmproj-BF16.gguf crashed the server on
+# load -- "mismatch between text model (n_embd = 5120) and mmproj
+# (n_embd = 2048)". It was built for a different, smaller model's vision
+# tower, not Qwen3.8-27B (n_embd 5120). Same applies to mmproj-F16.gguf
+# (same family of file, unverified for this model). Do not re-enable either
+# without a projector confirmed to report n_embd=5120 for its text side.
+# "none" is a real opt-out (added to the continuity script 2026-08-20); an
+# empty value is NOT equivalent -- it still falls through to auto-detect,
+# which is how this got picked up in the first place.
+LLAMA_MMPROJ=none
+
+LLAMA_HOST=0.0.0.0
+LLAMA_PORT=8080
+LLAMA_CTX_SIZE=262144
+LLAMA_PARALLEL=2
+LLAMA_THREADS=16
+LLAMA_GPU_LAYERS=99
+
+LLAMA_CACHE_TYPE_K=q4_0
+LLAMA_CACHE_TYPE_V=q4_0
+
+LLAMA_BATCH_SIZE=512
+LLAMA_UBATCH_SIZE=512
+
+# Native MTP speculative decoding (self-contained, no draft model file).
+# Measured 2026-08-20: 60-91% draft acceptance, stable under -np 2
+# concurrency and 5x-consecutive/multi-turn stress, no crashes/corruption.
+LLAMA_ENABLE_SPEC_DECODING=true
+LLAMA_SPEC_TYPE=draft-mtp
+LLAMA_DRAFT_MAX=2
+LLAMA_DRAFT_MIN=1
+LLAMA_DRAFT_P_MIN=0.75
+
+# Sampling requested for this profile (llama-server defaults are top-p 0.95,
+# top-k 40, temp 0.8 when unset -- see scripts/run-llama-server-continuity.sh).
+LLAMA_TEMP=0.7
+LLAMA_TOP_P=0.8
+LLAMA_TOP_K=20
+
+LLAMA_REPEAT_PENALTY=1.0
+# LLAMA_CACHE_REUSE stays unset here (carried over from the previous Qwen3.6
+# profile, where it never took effect). --cache-reuse (in-slot prompt PREFIX
+# reuse) is disabled at load time in server-context.cpp whenever
+# llama_kv_cache::get_can_shift() is false, which holds on three architectural
+# conditions, none of them a flag: arch==STEP35; n_pos_per_embd()>1 (M-RoPE);
+# or any layer with k_idx ("shifting would leave k_idx stale" -- the GDN
+# linear-attention layers). qwen35-arch models (3.6, 3.8) trip the latter.
+# The host-side prompt cache (--cache-ram) and --ctx-checkpoints are the
+# mechanisms that DO run here.
+LLAMA_SLOT_SAVE_PATH=/home/cogtek/.cache/inference/kv-slots
+LLAMA_LOG_FILE=/home/cogtek/llama.cpp/llama-server.log
+LLAMA_N_PREDICT=32768
+
+# CHAT TEMPLATE (2026-08-20): switched from qwen3.5-enhanced.jinja to
+# qwen-sharp.jinja -- peculiar-ragdoll/Qwen-Sharp-Chat-Templates
+# (template_version qwen3.8-froggeric-v22.1), vendored verbatim at
+# tools/agents/config/qwen-sharp.jinja. It targets Qwen 3.5/3.6/3.8 and is the
+# template for ALL qwen models here. Pinned artifact -- both are asserted in
+# test/chat-template-system-message.test.ts so a silent re-vendor fails CI:
+#   upstream revision  3dc34df52c63dd22ada21f96435e069deaa8d7da (2026-08-18)
+#   file sha256        d1f22a89eac3609dcfaa7b471b1f7d23bee2f084d275d26f4f8231d1d7908f4e
+#
+# WHAT IT FIXES: qwen3.5-enhanced.jinja hard-raised
+# 'No user query found in messages.' -- a server-side failure for the whole
+# request -- whenever no user turn survived its backwards scan, i.e. when every
+# user turn was a <tool_response> or when the transcript was system+assistant
+# only. Both shapes occur in normal agent loops. qwen-sharp renders them.
+#
+# WHAT STAYS COMPATIBLE (rendered and compared, not assumed):
+#   * the tool-call contract the model is told to EMIT is unchanged:
+#     <tool_call><function=NAME><parameter=K>V</parameter></function></tool_call>.
+#     anthropic_proxy.py's _HERMES_FUNCTION_RE / _TOOL_CALL_XML_RE parse it
+#     unchanged. The default is selected at render time by `_tool_format`, so
+#     the test asserts on RENDERED output -- the file contains a JSON emitter
+#     too, and a grep cannot tell which one is live.
+#   * the generation prompt matches the old template in BOTH enable_thinking
+#     states, so --reasoning auto --reasoning-format deepseek is unaffected.
+#   * a mid-conversation system message renders inline in both templates.
+#
+# WHAT ACTUALLY DIFFERS -- known and accepted:
+#   * HISTORY rendering of prior tool calls differs in whitespace only: with
+#     string-form `arguments` (the proxy's path -- it json.dumps them) the old
+#     template emitted "}\n</function>", the new one "}</function>"; and the
+#     separator between two tool calls in one assistant turn went "\n" ->
+#     "\n\n". Neither changes what the parsers accept.
+#   * the injected tool-calling instruction block was rewritten upstream: it now
+#     requires a <think> block before each tool call and forbids prose before
+#     the call. Behavioural, deliberate, not a wire change.
+#   * a system message containing an IMAGE now raises. Unreachable via the
+#     proxy, which flattens system to a text string before dispatch
+#     (anthropic_proxy.py ~4709). Direct-to-:8080 OpenAI clients can hit it.
+#   * unknown message roles no longer raise; they render as "[role]: content".
+#   * prior <think> blocks are retained across turns. Upstream sells this as a
+#     prompt-cache win; on the PROXY path it is a no-op, because the proxy
+#     already strips <think> from prior assistant turns on purpose
+#     (anthropic_proxy.py ~4725: "Stripping breaks the copy cycle"). On the
+#     direct path it re-introduces exactly that cycle -- pass
+#     chat_template_kwargs.preserve_thinking=false there if it shows up.
+#
+# chat_template_kwargs.reasoning_effort accepts
+# none|minimal|low|medium|high|xhigh|max. NOTE: 'none' forces thinking OFF,
+# overriding enable_thinking=true.
+# Do NOT switch this to "embedded" -- the model's own baked-in template still
+# carries the 'System message must be at the beginning.' raise.
+LLAMA_CHAT_TEMPLATE_FILE=/home/cogtek/dev/miller-tech/universal-agent-protocol/tools/agents/config/qwen-sharp.jinja
+
+# --cache-ram sizing inherited unchanged from the previous profile (measured
+# working set was for a different model, but the host-RAM cache mechanism
+# and its OOM history apply regardless of which model is loaded on the GPU
+# side). Re-measure with the eviction grep in qwen36-35b-a3b.env's history
+# if this model's prefill/eviction pattern differs materially.
+#
+# FLAG DRIFT: upstream renamed --checkpoint-every-n-tokens (old b8780
+# spelling) to --checkpoint-min-step. An unknown flag makes llama-server
+# exit 1, and systemd then burns its restart budget looping on it. Re-check
+# every flag here against `llama-server --help` whenever LLAMA_BIN moves.
+LLAMA_EXTRA_ARGS=--metrics --jinja --prio 2 --reasoning auto --reasoning-format deepseek --reasoning-budget 8000 --ctx-checkpoints 64 --checkpoint-min-step 2048 --cache-ram 16384

package/config/model-profiles/qwen35.json

@@ -82,10 +82,10 @@
 
     "llm_server": {
       "description": "Instance 1: Main LLM (Qwen3.5 35B A3B) — port 8080",
-      "command": "llama-server --model /path/to/Qwen3.5-35B-A3B-UD-IQ4_XS.gguf --chat-template-file chat_template.jinja --port 8080 --host 0.0.0.0 --ctx-size 131072 --gpu-layers 99 --cache-type-k q8_0 --cache-type-v q4_0 -fa on --threads 8 --batch-size 512 --ubatch-size 256 --mlock --metrics --n-predict 16384 --temp 0.3 --top-p 0.9 --top-k 20 --min-p 0.05",
+      "command": "llama-server --model /path/to/Qwen3.5-35B-A3B-UD-IQ4_XS.gguf --chat-template-file tools/agents/config/qwen-sharp.jinja --port 8080 --host 0.0.0.0 --ctx-size 131072 --gpu-layers 99 --cache-type-k q8_0 --cache-type-v q4_0 -fa on --threads 8 --batch-size 512 --ubatch-size 256 --mlock --metrics --n-predict 16384 --temp 0.3 --top-p 0.9 --top-k 20 --min-p 0.05",
       "flags": [
         "--model /path/to/Qwen3.5-35B-A3B-UD-IQ4_XS.gguf",
-        "--chat-template-file chat_template.jinja",
+        "--chat-template-file tools/agents/config/qwen-sharp.jinja",
         "--n-predict 16384",
         "--temp 0.3",
         "--top-p 0.9",

package/config/model-profiles/qwen36.json

@@ -79,7 +79,7 @@
       "description": "Instance 1: Main LLM (Qwen3.6 35B A3B) — port 8080, fronted by the proxy on :4000",
       "flags": [
         "--model /path/to/Qwen3.6-35B-A3B-UD-IQ4_XS.gguf",
-        "--chat-template-file chat_template.jinja",
+        "--chat-template-file tools/agents/config/qwen-sharp.jinja",
         "--n-predict 16384",
         "--temp 0.3",
         "--top-p 0.9",