2cc9370f3d
- vllm.service.j2: branch on role==embedding for --runner pooling --convert embed, --no-enable-prefix-caching, per-model trust_remote_code toggle (needed for nomic-embed-text-v1.5's custom NomicBertModel code), and enforce_eager toggle (needed to avoid CUDA graph capture OOM when co-resident with another vLLM process on this 24GB card). - tasks/verify.yml: split completions vs embedding smoke tests -- embedding-mode instances don't serve /v1/completions. Assert a non-empty embedding vector, not just HTTP 200. - host_vars/astro-orbiter: enable nomic-embed-text-v1.5 (port 8020), lower primary model's gpu_memory_utilization 0.95->0.90 + add enforce_eager after finding 0.95 crash-looped 6-7x before stabilizing with co-resident nomic-embed (real fix, confirmed via NRestarts=0 after clean stop/start, not luck). - Hindsight (values.yaml + externalsecret.yaml): cut LLM + embeddings over to vLLM (:8000, :8020), wire the previously-unset HINDSIGHT_API_EMBEDDINGS_* env vars for the first time, and swap the API key secret source from the Nous fallback item to vllm/api-key (vLLM enforces real auth, llama-swap did not). - README: document the embedding-mode branch, VRAM findings, and a genuine architecture gap -- vLLM's one-model-per-process design cannot replace llama-swap's 5-model LRU roster on this 24GB card, so 21 Hermes profiles' aux-model consumers (Qwen3-8B-no_think, Phi-3.5-mini, Meta-Llama-3.1-8B, Qwen2.5-Coder-14B) and OpenViking's VLM stay on llama-swap. Full teardown (t_6dff1ecc) needs a human decision on the aux-model strategy before it can proceed.
153 lines
8.4 KiB
YAML
153 lines
8.4 KiB
YAML
# ============================================================================
|
|
# Hindsight — helm values (Phase C). Consumed by the ArgoCD Application source 1
|
|
# via `helm.valueFiles: ["$values/values.yaml"]` (openviking multi-source pattern).
|
|
#
|
|
# Design decisions (all verified against chart v0.9.1 + rendered output):
|
|
# - Chart is the single source for the app (api, control-plane, services,
|
|
# probes, ingress). We do NOT hand-roll Deployments/Services.
|
|
# - Postgres is EXTERNAL (separate Deployment in this dir, firecrawl pattern)
|
|
# => postgresql.enabled: false, external.* points at hindsight-postgres:5432.
|
|
# - Secrets come from 1Password via ExternalSecret => existingSecret:
|
|
# hindsight-credentials. The chart does envFrom(secretRef) so
|
|
# HINDSIGHT_API_LLM_API_KEY / HINDSIGHT_API_MCP_AUTH_TOKEN are injected
|
|
# automatically; POSTGRES_PASSWORD is a secretKeyRef that K8s expands into
|
|
# HINDSIGHT_API_DATABASE_URL (verified with a live envFrom test pod).
|
|
# - LLM is the Nous free-tier inference endpoint
|
|
# (https://inference-api.nousresearch.com/v1), model
|
|
# `upstage/solar-pro4:free` (tool-calling, verified reflect). Fallback
|
|
# (documented, NOT deployed): `stepfun/step-3.7-flash:free`. API key via existingSecret
|
|
# envFrom (hindsight-credentials / HINDSIGHT_API_LLM_API_KEY), sourced from
|
|
# 1Password `nous` item per decision 4.
|
|
# - Ingress is driven through the chart's NATIVE ingress template (approved
|
|
# plan: "Ingress driven through values.yaml"). api.service.port=8888,
|
|
# controlPlane.service.port=3000.
|
|
# - Image tag defaults to .Values.version (root) when api.image.tag is unset,
|
|
# so version: "0.9.1" pins the API image to 0.9.1.
|
|
# ============================================================================
|
|
|
|
version: "0.9.1"
|
|
|
|
# ----------------------------------------------------------------------------
|
|
# External PostgreSQL (chart's bundled postgresql is disabled).
|
|
# password is the K8s env expansion `$(POSTGRES_PASSWORD)` — the chart defines
|
|
# POSTGRES_PASSWORD as a secretKeyRef (hindsight-credentials / postgres-password)
|
|
# earlier in the same container, so K8s substitutes it at container start.
|
|
# ----------------------------------------------------------------------------
|
|
postgresql:
|
|
enabled: false
|
|
external:
|
|
host: hindsight-postgres
|
|
port: 5432
|
|
username: hindsight
|
|
database: hindsight
|
|
password: $(POSTGRES_PASSWORD)
|
|
|
|
# ExternalSecret (from 1Password) that the chart injects via envFrom(secretRef).
|
|
# Keys it must expose: postgres-password, HINDSIGHT_API_LLM_API_KEY,
|
|
# HINDSIGHT_API_MCP_AUTH_TOKEN. See externalsecret.yaml in this dir.
|
|
existingSecret: hindsight-credentials
|
|
|
|
# ----------------------------------------------------------------------------
|
|
# API container environment (explicit env entries; the chart renders this map
|
|
# to individual env vars). LLM points at the Nous free-tier inference endpoint
|
|
# (https://inference-api.nousresearch.com/v1), model upstage/solar-pro4:free
|
|
# (tool-calling, verified retain/recall). stepfun/step-3.7-flash was swapped out
|
|
# because it rejects Hindsight's tagged fact-extraction (BadRequestError 400
|
|
# 'missing tags'); solar-pro4 is the previously-verified-good Nous model for the
|
|
# extract/retain path (tasks t_e3375410 / t_d0dffc3d).
|
|
# HINDSIGHT_API_LLM_API_KEY is NOT set here — it comes from the existingSecret
|
|
# via envFrom (1Password `nous` item).
|
|
# ----------------------------------------------------------------------------
|
|
api:
|
|
env:
|
|
# Cutover to vLLM (t_e6facb19, 2026-08-31): astro-orbiter's llama-swap is
|
|
# being retired for the completions+embeddings roles Hindsight needs.
|
|
# vLLM serves Qwen2.5-32B-Instruct-AWQ (:8000, completions) and
|
|
# nomic-embed-text-v1.5 (:8020, --runner pooling --convert embed) as two
|
|
# independent systemd-managed processes — validated end-to-end (health,
|
|
# /v1/chat/completions, /v1/embeddings returning a 768-dim vector) in a
|
|
# shadow window with llama-swap stopped. Model swap from Qwen3.8-27B to
|
|
# Qwen2.5-32B-Instruct-AWQ: both are capable instruction-tuned models:
|
|
# confirm structured-extraction/tagging behavior (the prior stepfun/solar
|
|
# 400 "missing tags" failure mode) holds up after this swap — see
|
|
# references/hindsight-backend-fallback-to-local-qwen-workflow.md if it
|
|
# regresses and llama-swap needs to come back temporarily.
|
|
HINDSIGHT_API_LLM_BASE_URL: "http://astro-orbiter:8000/v1"
|
|
HINDSIGHT_API_LLM_PROVIDER: "openai"
|
|
HINDSIGHT_API_LLM_MODEL: "Qwen2.5-32B-Instruct-AWQ"
|
|
# Embeddings: llama.cpp nomic-embed-text-v1.5 (llama-swap, :8001) ->
|
|
# vLLM nomic-embed-text-v1.5 (:8020). Same model weights, different
|
|
# serving stack. Hindsight's embeddings provider defaults to "local"
|
|
# (bundled sentence-transformers) unless overridden — explicit TEI/openai
|
|
# provider config was never wired for astro-orbiter previously (verified:
|
|
# `kubectl exec ... env | grep -i embed` showed NO HINDSIGHT_API_EMBEDDINGS_*
|
|
# vars set, confirming Hindsight was using its bundled local embedder, NOT
|
|
# astro-orbiter, before this change — despite the OpenViking-style stack
|
|
# documentation implying otherwise). Wiring the openai-compatible provider
|
|
# here for the first time to point Hindsight's actual embedding generation
|
|
# at the vLLM-served nomic-embed-text-v1.5, matching the model OpenViking
|
|
# already uses (consistency across mk-labs services).
|
|
HINDSIGHT_API_EMBEDDINGS_PROVIDER: "openai"
|
|
HINDSIGHT_API_EMBEDDINGS_OPENAI_BASE_URL: "http://astro-orbiter:8020/v1"
|
|
HINDSIGHT_API_EMBEDDINGS_OPENAI_MODEL: "nomic-embed-text-v1.5"
|
|
# --- t_d7f8cd65: fix 502s on the serial astro-orbiter node ---
|
|
# astro-orbiter is a single llama-swap process (serial: 1 generate at a
|
|
# time, ctx 64K). Hindsight's default LLM concurrency is 32, so a retain
|
|
# burst hits the node with N parallel calls -> the node rejects/times out
|
|
# the extras -> hindsight-api surfaces APITimeoutError as 502. astro-orbiter
|
|
# is the ONLY LLM endpoint (all ops route there), so cap the whole pool to
|
|
# 1 and pin retain to 1 as well. The upstream chart exposes these as native
|
|
# semaphore config (HINDSIGHT_API_*_MAX_CONCURRENT); no code change needed.
|
|
# Kept at 1 post-cutover: vLLM's single-process-per-model design is also
|
|
# effectively serial for a single Qwen2.5-32B instance under this GPU's
|
|
# VRAM budget (KV cache sized for low concurrency at max_model_len=8192).
|
|
HINDSIGHT_API_LLM_MAX_CONCURRENT: "1"
|
|
HINDSIGHT_API_RETAIN_LLM_MAX_CONCURRENT: "1"
|
|
# Client + per-request timeout. Default is 120s; a 29K-token retain runs
|
|
# ~29s and under load a single long retain can reach ~90s. Raise to 600s to
|
|
# cover the longest round-trip so the serial call never times out the client
|
|
# (keep >= ingress proxy-read-timeout below). Per-op retain timeout pins the
|
|
# retain path explicitly; the global timeout covers reflect/consolidation.
|
|
HINDSIGHT_API_LLM_TIMEOUT: "600"
|
|
HINDSIGHT_API_RETAIN_LLM_TIMEOUT: "600"
|
|
|
|
# ----------------------------------------------------------------------------
|
|
# Ingress via the chart's native template.
|
|
# /health,/v1,/mcp,/ext -> api:8888 (longest-prefix wins in nginx)
|
|
# / -> controlPlane:3000
|
|
# TLS secret hindsight-tls provisioned by the letsencrypt-prod issuer.
|
|
# ----------------------------------------------------------------------------
|
|
ingress:
|
|
enabled: true
|
|
className: "nginx"
|
|
annotations:
|
|
cert-manager.io/cluster-issuer: "letsencrypt-prod"
|
|
# Raised read/send timeout so a slow agentic reflect / long single retain
|
|
# (up to ~90s under load on the serial astro-orbiter node; client timeout
|
|
# is 600s per t_d7f8cd65) can complete before nginx cuts the connection.
|
|
# Raised 300 -> 600 (t_d7f8cd65) to cover the longest retain round-trip.
|
|
nginx.ingress.kubernetes.io/proxy-read-timeout: "600"
|
|
nginx.ingress.kubernetes.io/proxy-send-timeout: "600"
|
|
hosts:
|
|
- host: cosmic-rewind.local.mk-labs.cloud
|
|
paths:
|
|
- path: /health
|
|
pathType: Prefix
|
|
service: api
|
|
- path: /v1
|
|
pathType: Prefix
|
|
service: api
|
|
- path: /mcp
|
|
pathType: Prefix
|
|
service: api
|
|
- path: /ext
|
|
pathType: Prefix
|
|
service: api
|
|
- path: /
|
|
pathType: Prefix
|
|
service: controlPlane
|
|
tls:
|
|
- hosts:
|
|
- cosmic-rewind.local.mk-labs.cloud
|
|
secretName: hindsight-tls
|