Project_Velocity/.Agent Context/Desineuron AWS Coding Runtime Truth Book.md

# Desineuron AWS Coding Runtime Truth Book

Date: 2026-04-22  
Scope: Coding runtime, Roo Code access, NemoClaw runtime, ingress routing, GPU recovery, model staging

## 1. Current Runtime Truth

The Desineuron shared coding runtime has been cut over from Ollama to SGLang while preserving the public contracts already used by the team.

Locked production decisions:

- Public contract remains stable.
- GPU inference remains on the AWS GPU worker, not on the Linux-origin box.
- Linux-origin remains the control plane.
- Ingress remains the stable routed entrypoint.
- `Qwen 3.6 35B A3B` remains the production target model for the current `4 x L4` rollout.
- `NemoClaw` moves onto the same shared runtime.
- There is no production fallback to Ollama after cutover.

Current live public routes:

- `https://velocity.desineuron.in/llm`
- `https://llm.desineuron.in`

Current live API shape after cutover:

- `https://velocity.desineuron.in/llm/v1/models`
- `https://velocity.desineuron.in/llm/v1/chat/completions`
- `https://llm.desineuron.in/v1/models`
- `https://llm.desineuron.in/v1/chat/completions`
- GPU SGLang bind: `172.31.46.190:30100`
- Linux-origin LLM route-sync target port: `30100`

## 2. Infra Split

### Linux-origin

Responsibilities:

- owns route-sync logic
- owns operational orchestration
- updates ingress upstream target when GPU private IP changes
- does not host the heavy model runtime

### Ingress

Responsibilities:

- terminates public hostname
- renders stable reverse-proxy contracts
- forwards `/llm/*` and `llm.desineuron.in` to the current GPU target

### GPU worker

Responsibilities:

- hosts SGLang
- hosts model payloads on NVMe only
- serves Roo Code, Oracle runtime, runtime LLM, and NemoClaw inference

Non-negotiable rules:

- do not use the GPU public IP directly
- do not keep model state on root disk
- keep all large model/runtime caches on GPU NVMe

## 3. Live Hardware Target

Current worker class:

- `g6.12xlarge`
- `4 x NVIDIA L4`
- `96 GB VRAM total`

Serving profile for this hardware:

- tensor parallel size `4`
- prompt-prefix caching enabled
- async / continuous batching enabled through SGLang
- FlashInfer preferred where supported by the live CUDA stack

Measured validation on the live GPU worker:

- host class: `g6.12xlarge`
- GPU layout: `4 x NVIDIA L4`
- model path used for the validated runtime: `/opt/dlami/nvme/models/Qwen-Qwen3.6-35B-A3B-FP8`
- SGLang served model ID used for the test: `qwen3.6-35b-a3b`
- validated SGLang launch profile:
  - `--tp-size 4`
  - `--attention-backend flashinfer`
  - `--context-length 131072`
  - `--mem-fraction-static 0.88`
  - `--dist-init-addr 127.0.0.1:50000`
  - `--enable-metrics`
- required bind rule on this SGLang build:
  - public HTTP server must bind to the GPU private IP, not `0.0.0.0`
  - internal scheduler keeps a loopback listener on the API port
  - wildcard bind collides with that loopback listener on this build
- public validation after cutover:
  - `https://velocity.desineuron.in/llm/v1/models` returns `200`
  - `https://llm.desineuron.in/v1/models` returns `200`
  - streamed chat TTFT through public ingress measured at about `2.36 s`
  - one short non-stream completion measured about `33.86 completion tok/s`

## 4. Production Model Policy

### Primary production model

- user-facing family: `Qwen 3.6 35B A3B`
- exact SGLang served model ID: `qwen3.6-35b-a3b`

Why it remains live:

- fits the current `4 x L4` target
- already aligned with current team workflows
- suitable for coding/runtime use while the SGLang migration lands
- measured well enough for three concurrent coding users on the current hardware

### Staged future model on current L4 hardware

- `cyankiwi/Qwen3.5-122B-A10B-AWQ-4bit`

Status:

- acquisition/staging path is added
- not the live runtime on the current L4 cutover
- should be treated as a staged artifact for later runtime experimentation and hardware-fit validation

Why this is the right 122B staging path for the current worker:

- `4 x L4` is a better fit for an AWQ/int4 track than for an NVFP4 track
- this keeps the 122B experiment aligned with current hardware instead of assuming a Blackwell-oriented path

Why `txn545/Qwen3.5-122B-A10B-NVFP4` is not the active choice on L4:

- NVFP4 is not the safe default for the current L4 rollout
- if the team wants that track later, it should be treated as a separate hardware/runtime validation branch

Why no 122B model is the active live model in this round:

- the current migration is locked to preserving service continuity on the existing `4 x L4` worker
- the 122B track is a separate performance-fit and runtime-tuning exercise

## 5. Runtime Software Stack

Primary runtime after cutover:

- `SGLang`

Primary interface style:

- OpenAI-compatible `/v1/*`

Required runtime features:

- tensor parallel across all four GPUs
- prefix cache / prompt cache
- async scheduling
- continuous batching
- FlashInfer when supported by the live driver/runtime stack

Observed runtime note from the live bring-up:

- FlashInfer required `ninja-build` on the GPU box because it JIT-builds kernels on first run.
- The current GPU image needed:
  - `ninja-build`
  - `build-essential`
- After installing those packages, the FP8 runtime came up cleanly and served OpenAI-compatible traffic.

If stock SGLang underperforms:

- keep the same public routes
- tune CUDA/runtime behavior behind the same routed contract
- do not reintroduce Ollama fallback

## 6. Implemented Repo Changes

### Backend runtime service

File:

- `backend/services/runtime_llm_service.py`

Current state:

- provider catalog is standardized to `sglang`
- legacy provider names like `ollama` and `nemoclaw` are mapped into `sglang` to avoid immediate caller breakage
- model discovery uses `/v1/models`

### NemoClaw client

File:

- `backend/services/nemoclaw_client.py`

Current state:

- production path now targets the shared SGLang/OpenAI-compatible endpoint
- NVIDIA and Ollama production fallback logic is removed from the runtime path
- legacy env names still seed config where needed

### Prompt expander

File:

- `comfy_engine/scripts/prompt_expander.py`

Current state:

- now uses the shared OpenAI-compatible runtime instead of Ollama `/api/generate`

### NemoClaw deploy helper

File:

- `backend/scripts/nemoclaw_deploy.sh`

Current state:

- rewritten around SGLang-compatible inference
- no Ollama-era deployment assumptions

## 7. Route Sync And Stable Hostnames

Route-sync files:

- `infrastructure/desineuron_ingress/sync_llm_route.py`
- `infrastructure/desineuron_ingress/run_llm_route_sync.sh`
- `infrastructure/desineuron_ingress/desineuron-llm-route-sync.service`
- `infrastructure/desineuron_ingress/desineuron-llm-route-sync.timer`
- `infrastructure/desineuron_ingress/install_linux_llm_route_sync.sh`

Important behavior:

- Linux-origin discovers the current GPU private IP
- Linux-origin updates ingress-managed route state
- ingress forwards `llm.desineuron.in` and `/llm/*` to the GPU worker

Current safe default route-sync port in the repo:

- `11434`

Reason:

- the repo now contains the SGLang installer and watchdog, but the public route should not auto-cut from Ollama to SGLang until the GPU runtime is actually installed and validated on-host
- when SGLang is installed on the GPU worker, operators should flip `LLM_ROUTE_PORT` to the live SGLang port and then run route-sync

Manual operator-safe route sync entrypoint:

- `/usr/local/bin/run_llm_route_sync.sh`

This avoids the prior failure mode where operators accidentally used a system Python without `boto3`.

## 8. GPU Watchdog And Auto-Recovery

Added GPU-side scripts:

- `infrastructure/desineuron_ingress/install_gpu_sglang_runtime.sh`
- `infrastructure/desineuron_ingress/install_gpu_sglang_watchdog.sh`

Installed unit names expected on the GPU worker:

- `desineuron-sglang.service`
- `desineuron-sglang-watchdog.service`
- `desineuron-sglang-watchdog.timer`

Recovery policy:

- ensure the SGLang service is running
- verify `/v1/models` health locally
- if the configured model path is missing, rehydrate from the canonical source
- only report healthy after successful verification

Required recovery assertions for the SGLang watchdog:

- confirm the process is serving `/v1/models`
- confirm the returned model list contains `qwen3.6-35b-a3b`
- confirm all 4 GPUs are engaged during model load
- confirm FlashInfer dependencies are present before declaring runtime healthy

## 9. Model Rehydration And Staging

Added staging helper:

- `infrastructure/desineuron_ingress/acquire_qwen35_122b_nvfp4.sh`

Purpose:

- stages `cyankiwi/Qwen3.5-122B-A10B-AWQ-4bit` onto GPU NVMe by default
- does not automatically flip production traffic to that model

Expected current live model path style:

- `/opt/dlami/nvme/models/Qwen-Qwen3.6-35B-A3B-FP8`

Expected staged 122B path style:

- `/opt/dlami/nvme/models/cyankiwi-Qwen3.5-122B-A10B-AWQ-4bit`

## 10. Roo Code Team Setup

After SGLang cutover, team members should stop using the Ollama provider mode for Desineuron-hosted inference.

Canonical team profile:

- API Provider: OpenAI-compatible / custom OpenAI
- Base URL: `https://llm.desineuron.in/v1`
- Model: `qwen3.6-35b-a3b`
- Temperature: `0.1` to `0.2`
- Server context ceiling: `131072`
- Recommended Roo context: `131072`

Team decision for this wave:

- all three team members can target `128K` context through the same shared runtime
- if real concurrent repo-heavy usage causes OOM or latency regression, the first rollback knob is the client context setting, not the model family
- the current production-ready long-context path is pure VRAM on `4 x L4`, not host-RAM spill

## 11. Measured SGLang Performance

Benchmark date:

- `2026-04-22`

Benchmark topology:

- live AWS GPU worker
- `SGLang + Qwen 3.6 35B A3B FP8`
- tensor parallel `4`
- FlashInfer enabled
- async scheduler / SGLang default continuous batching path
- prompt-prefix caching available in runtime
- server context ceiling: `131072`

Measured results:

- time to first token: `0.12 s`
- streamed completion wall time for a short coding/planning answer: `1.31 s`
- test concurrency: `3`
- aggregate wall time for `3 x 256-token` responses: `3.61 s`
- aggregate completion tokens: `768`
- aggregate prompt tokens: `168`
- aggregate total tokens: `936`
- aggregate completion throughput: `212.76 tokens/s`

Per-request timing under `3` concurrent requests:

- request 1: `3.608 s` for `256` completion tokens
- request 2: `3.609 s` for `256` completion tokens
- request 3: `3.608 s` for `256` completion tokens

Long-context smoke validation:

- prompt size validated: `50010` prompt tokens
- completion size: `8` tokens
- total request size: `50018` tokens
- wall time: `8.345 s`

Operational interpretation:

- the runtime is fast enough for three simultaneous coding users
- TTFT is already in the sub-200 ms range on the warmed runtime
- aggregate decode throughput is materially better than the previous Ollama-backed path while holding a `128K` server context ceiling
- `Qwen 3.6 35B A3B` is the correct production choice for the current one-week delivery window

## 12. Cutover Guidance

Use this model ID consistently across SGLang-facing clients:

- `qwen3.6-35b-a3b`

Do not use this older Ollama-style model ID against SGLang:

- `qwen3.6:35b-a3b`

Why:

- SGLang rejects colons in `served_model_name`
- the colon is reserved internally for adapter syntax

Backend compatibility note:

- the Velocity backend can still map legacy provider naming internally
- external Roo Code and OpenAI-compatible clients should use the hyphenated SGLang model ID only

Canonical Roo configuration:

- API Provider: `OpenAI-compatible` or `Custom OpenAI`
- Base URL: `https://llm.desineuron.in/v1`
- Model: `qwen3.6-35b-a3b`
- Context window: `131072`
- Temperature: `0.1` to `0.2`

Recommended initial values:

- `Base URL`: `https://llm.desineuron.in/v1`
- `Model`: `qwen3.6-35b-a3b`
- `Context Window Size (num_ctx equivalent)`: `131072`

Do not use:

- Ollama provider mode pointing at the public Desineuron route after the cutover

Reason:

- the stable contract is moving to SGLang's OpenAI-compatible interface

## 13. Most Efficient Working Long-Context Strategy On Current Hardware

Strategies tested against the live `4 x L4` worker:

1. Pure-VRAM `131072` context on SGLang with tensor parallel `4`
Result:

- works
- preserves sub-200 ms TTFT on warm short prompts
- preserved about `212.76 tok/s` aggregate completion throughput in the 3-user benchmark

2. Hierarchical host-memory cache with `131072` context
Result:

- not production-safe on the current stack for this model
- first failed on a model-specific `page_size=1` requirement for the hybrid Mamba cache
- second attempt progressed further but one rank died with exit code `-9`
- current interpretation: this path is materially less stable than the pure-VRAM profile

Current decision:

- keep `131072` in VRAM as the production target
- do not use host-RAM hierarchical cache for this model in the current rollout
- if more headroom is needed later, tune kernels and scheduling first before re-opening host-memory spill

## 14. NemoClaw Runtime Policy

NemoClaw should use the same shared SGLang runtime as:

- Roo Code
- Oracle runtime
- backend runtime LLM jobs

This is a deliberate single-stack decision:

- one serving runtime
- one model family for the current wave
- one stable routed contract

If later profiles differ, express that with config, not with a second serving stack in this phase.

## 15. Endpoint Checklist

These should work after cutover:

- `https://velocity.desineuron.in/llm/v1/models`
- `https://velocity.desineuron.in/llm/v1/chat/completions`
- `https://llm.desineuron.in/v1/models`
- `https://llm.desineuron.in/v1/chat/completions`

Internal backend envs:

- `LLM_BASE_URL`
- `SGLANG_BASE_URL`
- `SGLANG_CHAT_URL`
- `SGLANG_MODELS_URL`
- `SGLANG_MODEL`
- `SGLANG_API_TOKEN`

## 16. What Is Left

Still required to complete the migration end to end:

1. Persist the `131072` launch profile into the GPU systemd runtime using the updated installer.
2. Reinstall or update the GPU watchdog so it validates the same `131072` service profile.
3. Repoint Linux-origin route-sync env from `11434` to the live SGLang port after GPU validation.
4. Validate both public routes against `/v1/models`.
5. Run one more public-route benchmark through ingress after cutover to capture real routed TTFT.
6. Generate tuned L4-specific runtime configs if we want to push further on throughput without lowering context.
7. Keep the 122B track separate; it is not part of the current production coding-runtime choice.

## 17. Team Hand-Off

For Roo Code today, once cutover is complete, the team only needs:

- Base URL: `https://llm.desineuron.in/v1`
- Model: `qwen3.6-35b-a3b`
- Context window: `131072`
- Provider type: OpenAI-compatible

For operators, the important truth is:

- Linux-origin controls routing
- ingress owns the stable hostname
- GPU box owns inference
- NVMe owns model state
- SGLang is the production runtime