@miller-tech/uap 1.39.0 → 1.40.1

package/docs/benchmarks/SPECULATIVE_DECODING_JOURNEY_2026-03.md

@@ -1,221 +0,0 @@
-# Speculative Decoding Journey (2026-03)
-
-This document records the end-to-end speculative decoding stabilization journey across `llama.cpp` runtime tuning and `uap-anthropic-proxy` guardrails, including fixes, benchmark results, and the production profile now in use.
-
-## Scope
-
-- Runtime: `llama.cpp` with Qwen3.5 models, CUDA, `ctx-size=262144`.
-- Gateway: Anthropic-compatible proxy (`tools/agents/scripts/anthropic_proxy.py`).
-- Client behavior: agentic coding loops with tool calls (Claude Code style).
-
-## Goals
-
-1. Preserve high speculative decoding throughput.
-2. Eliminate pathological loops and malformed visible output.
-3. Keep tool-call behavior reliable under long sessions.
-4. Keep production context window at `262144`.
-
-## Phase 1 - Llama.cpp Speculative Stability
-
-### Problems Observed
-
-- Rollback loops and instability under aggressive speculative settings.
-- `find_slot` and related server warnings during long agentic sessions.
-- Throughput regressions compared to known fast baseline.
-
-### Work Performed
-
-- Implemented and tested multiple rollback strategies in `llama.cpp` worktree branches.
-- Compared baseline fast commit vs newer speculative logic.
-- Restored proven fast runtime path for production service while preserving learned guardrails.
-
-### Key Runtime Decisions
-
-- Keep production on fast validated binary lineage (`029edcafc` baseline family).
-- Use strict balanced speculative profile for 35B operations:
-  - `speculative.n_max=12`
-  - `speculative.n_min=2`
-  - `speculative.p_min=0.80`
-
-### Representative Throughput Findings
-
-- Qwen3.5-27B, `ctx=262144`, q4 KV cache:
-  - No spec: ~43 tok/s coding, ~41 tok/s pattern.
-  - Spec (balanced): ~43 tok/s coding, ~102 tok/s pattern.
-  - Main uplift appears in pattern-heavy turns, not all coding turns.
-
-## Phase 2 - Proxy Reasoning Fallback Leak Fix
-
-### Problems Observed
-
-- Empty visible output (`output_tokens=0`) with large hidden reasoning payloads.
-- Proxy emitted malformed chain-of-thought text as fallback, causing user-visible garbage:
-  - repeated fragments like `</parameter>`, tool schema echoes, policy text loops.
-
-### Fixes Implemented
-
-- Added explicit streaming fallback policy:
-  - `PROXY_STREAM_REASONING_FALLBACK=off|sanitized|visible`
-  - `PROXY_STREAM_REASONING_MAX_CHARS`
-- Set production default to `off`.
-
-### Result
-
-- Malformed reasoning fallback leakage is suppressed by default.
-- Debugging remains possible with `sanitized`/`visible` modes when intentionally enabled.
-
-## Phase 3 - Token Floor and Prune Controls
-
-### Problems Observed
-
-- Hardcoded `max_tokens` floor (`16384`) forced very long failure turns.
-- Pruning threshold flag alone could trigger pruning path without meaningful message reduction.
-
-### Fixes Implemented
-
-- Added configurable max token floor:
-  - `PROXY_MAX_TOKENS_FLOOR` (`0` disables floor)
-- Added configurable prune target:
-  - `PROXY_CONTEXT_PRUNE_TARGET_FRACTION`
-
-### Live A/B Result (Production-Like)
-
-`PROXY_MAX_TOKENS_FLOOR=16384` vs `4096`:
-
-- Silent reasoning-heavy turn:
-  - `16384`: avg `78.749s`
-  - `4096`: avg `19.777s`
-  - Latency reduction: ~`74.9%`
-  - Predicted throughput unchanged (~`208 tok/s` class)
-- Normal tool turns remained stable and slightly faster with `4096`.
-
-## Phase 4 - Malformed Tool-Loop Hardening
-
-### Problem Pattern
-
-Under adversarial or degraded prompt states, the model can emit pseudo-tool text instead of valid tool calls, e.g.:
-
-- `</parameter>` fragments
-- echoed policy snippets (`you MUST call a tool...`)
-- long no-progress text with no `tool_calls`
-
-### Feature Set Added (Flag Controlled)
-
-1. **Malformed tool guardrail + retry**
-   - `PROXY_MALFORMED_TOOL_GUARDRAIL`
-   - `PROXY_MALFORMED_TOOL_RETRY_MAX`
-   - `PROXY_MALFORMED_TOOL_RETRY_MAX_TOKENS`
-   - `PROXY_MALFORMED_TOOL_RETRY_TEMPERATURE`
-
-2. **Strict stream guardrail path**
-   - `PROXY_MALFORMED_TOOL_STREAM_STRICT`
-   - For stream+tools requests, proxy runs guarded non-stream upstream call, then replays SSE.
-
-3. **Tool narrowing (optional)**
-   - `PROXY_TOOL_NARROWING`
-   - `PROXY_TOOL_NARROWING_KEEP`
-   - `PROXY_TOOL_NARROWING_MIN_TOOLS`
-
-4. **Disable thinking on tool turns (optional)**
-   - `PROXY_DISABLE_THINKING_ON_TOOL_TURNS`
-
-5. **Session contamination breaker (optional safety net)**
-   - `PROXY_SESSION_CONTAMINATION_BREAKER`
-   - `PROXY_SESSION_CONTAMINATION_THRESHOLD`
-   - `PROXY_SESSION_CONTAMINATION_KEEP_LAST`
-
-6. **Agentic supplement mode**
-   - `PROXY_AGENTIC_SUPPLEMENT_MODE=clean|legacy`
-
-### Test Coverage
-
-- Unit tests in `tools/agents/tests/test_anthropic_proxy_streaming.py`
-- Current targeted suite count in this workstream: `16` passing tests.
-
-## Benchmark Highlights (Per-Option Toggles)
-
-### Artifact Stress Benchmark (v3)
-
-Source: `/tmp/proxy_visibility_benchmark_v3.json`
-
-| Mode | Key Flags | Outcome Summary |
-| --- | --- | --- |
-| Baseline | none | no tool call, policy-echo text surfaced |
-| Option 1 | malformed guardrail + strict stream | malformed detected and retried; returned `tool_use` with empty visible text |
-| Option 2 | tool narrowing only | not sufficient alone in stress case |
-| Option 3 | disable thinking only | not sufficient alone in stress case |
-| Option 4 | contamination breaker only | not sufficient alone in this synthetic workload |
-| Option 5 | clean supplement only | not sufficient alone in stress case |
-
-### Practical Conclusion
-
-- Strongest primary mitigation: **Option 1** (malformed guardrail + strict stream + bounded retry).
-- Other options are secondary tuning aids and should not replace Option 1 for this failure class.
-
-## 10-Turn Live Stability Soak
-
-Source: `/tmp/proxy_10turn_soak_results.json`
-
-- 10 turns, alternating malformed-stress and normal tool-call turns, single live session id.
-- Results:
-  - Error rate: `0.0%`
-  - Malformed visible output rate (stress turns): `0.0%`
-  - Normal tool-call success rate: `100.0%`
-  - Duration p50/p95: `10.2s` / `21.366s`
-  - Stop reasons: `tool_use=6`, `max_tokens=3`, `end_turn=1`
-
-## Production Profile (Current)
-
-File: `/home/cogtek/.config/uap/anthropic-proxy.env`
-
-```bash
-PROXY_MAX_TOKENS_FLOOR=4096
-PROXY_STREAM_REASONING_FALLBACK=off
-
-PROXY_MALFORMED_TOOL_GUARDRAIL=on
-PROXY_MALFORMED_TOOL_STREAM_STRICT=on
-PROXY_MALFORMED_TOOL_RETRY_MAX=1
-PROXY_MALFORMED_TOOL_RETRY_MAX_TOKENS=512
-PROXY_MALFORMED_TOOL_RETRY_TEMPERATURE=0
-
-PROXY_TOOL_NARROWING=off
-PROXY_DISABLE_THINKING_ON_TOOL_TURNS=off
-PROXY_SESSION_CONTAMINATION_BREAKER=off
-PROXY_AGENTIC_SUPPLEMENT_MODE=legacy
-```
-
-Rationale:
-
-- Keep the strongest practical fix enabled (malformed guardrail + strict stream path).
-- Keep latency-optimized floor (`4096`).
-- Keep optional secondary heuristics off unless new evidence warrants enablement.
-
-## Reproduction Checklist
-
-1. Restart services:
-
-```bash
-systemctl --user restart uap-llama-server.service
-systemctl --user restart uap-anthropic-proxy.service
-```
-
-2. Run targeted unit tests:
-
-```bash
-python3 -m pytest tools/agents/tests/test_anthropic_proxy_streaming.py -q
-```
-
-3. Run soak script (or equivalent alternating malformed/normal stream sequence).
-
-4. Validate logs:
-
-- `MALFORMED TOOL PAYLOAD`
-- `MALFORMED RETRY ...`
-- `STRICT STREAM GUARDRAIL`
-- Absence of user-visible malformed fragments.
-
-## Open Follow-Ups
-
-- Add a dedicated persistent benchmark harness under `scripts/` for this exact soak profile.
-- Add branch/commit links from `llama.cpp` worktrees for cross-repo traceability.
-- Optionally evaluate enabling `PROXY_TOOL_NARROWING` in production only after longer mixed-workload soak data.