deploy-vllm: add embedding-mode support, cut over Hindsight to vLLM (t_e6facb19)

- 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.

ansible/host_vars/astro-orbiter/vars.yml

@@ -109,3 +109,57 @@ llm_staged_models:
     size_bytes: 5027784224
     source_repo: "bartowski/Qwen_Qwen3-8B-GGUF"
 
+# --- deploy-vllm role: vllm_models override (t_e6facb19, 2026-08-31) --------
+# Ansible's hash_behaviour is "replace" (see ansible.cfg) — a host_vars list
+# variable REPLACES the role default list wholesale, it does not deep-merge.
+# This is therefore a full copy of roles/deploy-vllm/defaults/main.yml's
+# vllm_models with ONE change: nomic-embed-text-v1.5.enabled flipped to true,
+# now that vllm.service.j2 has an embedding-mode branch (--runner pooling
+# --convert embed --trust-remote-code) tested end-to-end in a shadow window.
+# Primary (Qwen2.5-32B-Instruct-AWQ) and aux (Qwen3-8B-AWQ) entries are
+# unchanged from role defaults — reproduced here only because the whole list
+# must be redefined together. Keep this in sync with defaults/main.yml if the
+# role's model roster changes.
+vllm_models:
+  - id: "Qwen2.5-32B-Instruct-AWQ"
+    hf_repo: "Qwen/Qwen2.5-32B-Instruct-AWQ"
+    role: primary
+    quantization: awq
+    port: 8000
+    max_model_len: 8192
+    # 0.95 (role default) OOM'd during CUDA graph capture once nomic-embed
+    # (role: embedding, ~814MiB actual, not the nominal 300MB) is co-resident
+    # on the same 24GB card (t_e6facb19, 2026-08-31): KV cache allocation
+    # succeeded (14,720 tokens) but graph capture needed ~20MiB more than the
+    # 0.95 budget left after nomic's share. Two independent, permanent
+    # co-residents (unlike t_ca1af9fb's shadow-window test, which had the
+    # whole 24GB free) need either a lower utilization ceiling or no graph
+    # capture. enforce_eager avoids the whole cudagraph capture memory spike
+    # entirely — small throughput cost, no OOM risk, safer for a fixed
+    # multi-process VRAM budget than tuning utilization percentages by hand.
+    # Even WITH enforce_eager, 0.95 left only ~847MiB genuinely free out of
+    # 24576MiB total (23,729MiB used) and both services crash-looped 6-7x
+    # during warmup/KV-cache sizing before stabilizing — too fragile for a
+    # permanent two-process co-residency. Lowered to 0.90 for real headroom
+    # (~1.6GiB free), confirmed clean single-attempt start with no retries.
+    gpu_memory_utilization: 0.90
+    enforce_eager: true
+    enabled: true
+  - id: "Qwen3-8B-AWQ"
+    hf_repo: "Qwen/Qwen3-8B-AWQ"
+    role: aux
+    quantization: awq
+    port: 8010
+    max_model_len: 32768
+    gpu_memory_utilization: 0.15
+    enabled: false
+  - id: "nomic-embed-text-v1.5"
+    hf_repo: "nomic-ai/nomic-embed-text-v1.5"
+    role: embedding
+    quantization: none
+    port: 8020
+    max_model_len: 2048
+    gpu_memory_utilization: 0.05
+    trust_remote_code: true
+    enabled: true
+

ansible/roles/deploy-vllm/README.md

@@ -82,11 +82,6 @@ supports (`--quantization awq`) and which fit the VRAM budget:
 
 ## Known Gaps / Follow-ups
 
-- The `nomic-embed-text-v1.5` entry in `vllm_models` is present but the
-  `vllm.service.j2` template does not yet branch for embedding-mode flags
-  (`--task embed`). Do not flip `enabled: true` on it without first adding
-  that branch and testing `/v1/embeddings` — this is what Hindsight retain
-  actually depends on, so get it right before cutover.
 - Quarterly API key rotation is documented (`/etc/vllm/API_KEY_ROTATION.md`
   on the target, rendered by `tasks/api-key.yml`) but not automated — no cron
   job exists to force rotation on a schedule. Consider a follow-up cron task
@@ -94,6 +89,106 @@ supports (`--quantization awq`) and which fit the VRAM budget:
 - `vllm_service_enabled` defaults to `false` deliberately — see "Deliberate
   staging-first default" above. Flip together with the cutover step, not
   before.
+- **vLLM cannot replace llama-swap's full model roster on this card (t_e6facb19, 2026-08-31).**
+  This role only wires two of llama-swap's five served models: the primary
+  completions model (Qwen2.5-32B-Instruct-AWQ, replacing Qwen3.8-27B) and
+  the embedding model (nomic-embed-text-v1.5). llama-swap ALSO serves
+  Qwen3-8B-Q4_K_M(-no_think), Phi-3.5-mini-instruct-Q8_0,
+  Meta-Llama-3.1-8B-Instruct-Q4_K_M, and Qwen2.5-Coder-14B-Instruct-Q4_K_M —
+  21 Hermes agent profiles' `custom_providers` reference these model IDs for
+  aux tasks (skills_hub, approval, mcp, title_generation, profile_describer,
+  compression). vLLM 0.28 serves **one model per process**; running 5-6
+  separate vLLM processes concurrently does not fit a 24GB card (each
+  process reserves its own CUDA context + weights + KV cache, unlike
+  llama-swap's matrix DSL which time-shares one GPU across LRU-evicted
+  processes). **Full llama-swap teardown (t_6dff1ecc) cannot proceed until
+  either:** (a) the aux-model consumers are migrated to a different backend
+  (Anthropic, or a smaller local llama.cpp router kept alongside vLLM), or
+  (b) vLLM gains a comparable multi-model time-sharing mode. This is a
+  genuine architecture gap, not a missing role feature — flagging for a
+  human decision on the aux-model strategy before teardown can be
+  unconditionally safe.
+
+## Embedding-mode support (t_e6facb19, 2026-08-31)
+
+`vllm.service.j2` now branches on `role: embedding` entries in `vllm_models`:
+adds `--runner pooling --convert embed` (vLLM's embedding-serving flags —
+see https://docs.vllm.ai/en/latest/models/pooling_models/embed/) and
+`--no-enable-prefix-caching` (prefix caching is a completions-only
+optimization; irrelevant and safely disabled for pooling). An additional
+per-model `trust_remote_code: true` toggle renders `--trust-remote-code`
+when set — required for `nomic-ai/nomic-embed-text-v1.5`, which ships
+custom `NomicBertModel` modeling code on its HF repo.
+
+**Verification does NOT run `/v1/completions` against embedding-mode
+instances** (they don't serve that endpoint — a completions request 400s
+immediately). `tasks/verify.yml` splits `vllm_enabled_models` by `role` and
+runs the appropriate smoke test per group: completions models get the
+`/v1/completions` "capital of France" test; embedding models get a real
+`/v1/embeddings` POST with an `ansible.builtin.assert` on a non-empty
+`data[0].embedding` array (not just HTTP 200 — an empty/malformed vector
+would still 200).
+
+**Critical VRAM finding: co-resident completions + embedding vLLM processes
+need MORE headroom than either alone, and CUDA graph capture is the failure
+mode, not KV cache sizing.** Enabling `nomic-embed-text-v1.5` alongside the
+primary Qwen2.5-32B model at the role-default `gpu_memory_utilization: 0.95`
+crash-looped repeatedly:
+- First failure: `torch.OutOfMemoryError` during `capture_model()` (CUDA
+  graph capture) — KV cache sizing itself succeeded (14,720 tokens
+  allocated), but graph capture needed ~20MiB more than the 0.95 budget had
+  left once nomic's embedding process (814MiB actual, not the nominal
+  ~300MB estimate in the model roster table) claimed its share.
+- Fix attempt 1: added a per-model `enforce_eager: true` template branch
+  (`--enforce-eager` skips CUDA graph capture entirely) — this stopped the
+  graph-capture OOM but the combined processes still landed at only
+  ~847MiB genuinely free out of 24,576MiB, and both services crash-looped
+  6-7 times during warmup before finally stabilizing (each attempt leaves
+  transient VRAM that the next attempt fights over, extending time-to-stable
+  well past a single health-check retry window).
+- Fix attempt 2 (final, verified stable): lowered the primary model's
+  `gpu_memory_utilization` from 0.95 to **0.90** (host_vars override) in
+  addition to `enforce_eager: true`. Result: clean single-attempt start for
+  both services, `NRestarts=0`, ~2GB genuinely free (22,577MiB used /
+  24,576MiB total). Confirmed via `systemctl show <unit> -p NRestarts` after
+  a full stop/start cycle — 0.95 was NOT a fluke of Restart=always masking
+  the underlying fragility; 0.90 is a real, reproducible fix.
+- **Takeaway for future multi-process vLLM VRAM budgeting on this host:**
+  do not just check "does it eventually come up" — check `NRestarts` and
+  free VRAM headroom after a clean stop/start. A model that "works" after
+  6 crash-loop retries is not production-stable; the retries themselves are
+  evidence the utilization ceiling is too tight for the actual (not
+  nominal) footprint of co-resident processes.
+
+## Consumer cutover status (t_e6facb19, 2026-08-31)
+
+**Cut over (validated end-to-end):**
+- Hindsight (`cluster/applications/hindsight/values.yaml` +
+  `externalsecret.yaml`): `HINDSIGHT_API_LLM_BASE_URL` → vLLM `:8000`
+  (Qwen2.5-32B-Instruct-AWQ), plus newly-wired `HINDSIGHT_API_EMBEDDINGS_*`
+  env vars pointing at vLLM `:8020` (nomic-embed-text-v1.5). Both endpoints
+  require vLLM's real API key (unlike llama-swap, which accepted
+  any/no key) — ExternalSecret now reads `op://mk-labs/vllm/api-key`
+  (item "vllm") instead of the prior Nous fallback item, and reuses the same
+  key value for `HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY` (both vLLM
+  endpoints share one key file per `tasks/api-key.yml`).
+
+**NOT cut over — genuine scope gap requiring a human decision, see
+"Known Gaps" above:** the 21 Hermes agent profiles' aux-model
+`custom_providers` entries (Qwen3-8B-no_think, Phi-3.5-mini, Meta-Llama-3.1-8B,
+Qwen2.5-Coder-14B) still point at llama-swap `:8001` — vLLM has no
+equivalent multi-model serving mode on this 24GB card. llama-swap MUST stay
+up to serve these until that gap is resolved. This is why the teardown task
+(t_6dff1ecc) remains blocked even after this task's completion — see the
+comment posted there.
+
+- OpenViking (`cluster/platform/openviking/values.yaml`): still points at
+  llama-swap `:8001` (`Meta-Llama-3.1-8B-Instruct-Q4_K_M` VLM + nomic-embed
+  for dense embeddings). Left unchanged — its VLM model has no vLLM
+  equivalent staged, and migrating only its embedding path while leaving its
+  VLM on llama-swap would still require llama-swap up, providing zero
+  teardown benefit. Flagged, not touched, per the same aux-model gap above.
+
 
 ## Validation Log (2026-08-31, t_ca1af9fb)
 

ansible/roles/deploy-vllm/defaults/main.yml

@@ -55,6 +55,10 @@ vllm_models:
     port: 8020
     max_model_len: 2048
     gpu_memory_utilization: 0.05
+    # NomicBertModel ships custom modeling code on the HF repo (rotary/ALiBi
+    # variant) — vLLM needs --trust-remote-code to load it, same requirement
+    # as sentence-transformers/llama.cpp. Wired into vllm.service.j2 (t_e6facb19).
+    trust_remote_code: true
     enabled: false
 
 # --- systemd ---------------------------------------------------------------

ansible/roles/deploy-vllm/tasks/verify.yml

@@ -71,7 +71,12 @@
   loop_control:
     label: "{{ item.item.id }}"
 
-- name: Run a live completion smoke test against each enabled instance
+- name: Split enabled models into completion-serving vs embedding for the right smoke test
+  ansible.builtin.set_fact:
+    vllm_completion_models: "{{ vllm_enabled_models | rejectattr('role', 'equalto', 'embedding') | list }}"
+    vllm_embedding_models: "{{ vllm_enabled_models | selectattr('role', 'equalto', 'embedding') | list }}"
+
+- name: Run a live completion smoke test against each completion-serving instance
   ansible.builtin.uri:
     url: "http://127.0.0.1:{{ item.port }}/v1/completions"
     method: POST
@@ -86,7 +91,7 @@
       temperature: 0
     timeout: 60
     status_code: 200
-  loop: "{{ vllm_enabled_models }}"
+  loop: "{{ vllm_completion_models }}"
   loop_control:
     label: "{{ item.id }}"
   register: vllm_completion_test
@@ -99,12 +104,50 @@
   loop_control:
     label: "{{ item.item.id }}"
 
-- name: Check journalctl for the primary unit is free of ERROR/Traceback since last start
+# Embedding-mode vLLM instances (--runner pooling --convert embed) do NOT
+# serve /v1/completions — only /v1/embeddings (and /pooling). A completions
+# smoke test against one 400s immediately. Verify with a real vector request
+# instead, and assert the response actually contains a non-empty float vector
+# (not just HTTP 200 — an empty/malformed embedding would still 200).
+- name: Run a live embeddings smoke test against each embedding-mode instance
+  ansible.builtin.uri:
+    url: "http://127.0.0.1:{{ item.port }}/v1/embeddings"
+    method: POST
+    headers:
+      Authorization: "Bearer {{ vllm_api_key_lookup.stdout }}"
+      Content-Type: "application/json"
+    body_format: json
+    body:
+      model: "{{ item.id }}"
+      input: "The capital of France is Paris."
+    timeout: 60
+    status_code: 200
+    return_content: true
+  loop: "{{ vllm_embedding_models }}"
+  loop_control:
+    label: "{{ item.id }}"
+  register: vllm_embedding_test
+  no_log: true
+
+- name: Assert embeddings smoke test returned a non-empty float vector
+  ansible.builtin.assert:
+    that:
+      - (item.json.data[0].embedding | length) > 0
+    fail_msg: "/v1/embeddings on port {{ item.item.port }} did not return a non-empty embedding vector"
+    success_msg: "/v1/embeddings confirmed {{ item.item.id }} returns a {{ item.json.data[0].embedding | length }}-dim vector"
+  loop: "{{ vllm_embedding_test.results }}"
+  loop_control:
+    label: "{{ item.item.id }}"
+
+- name: Check journalctl for each enabled unit is free of ERROR/Traceback since last start
   ansible.builtin.shell: |
     set -o pipefail
-    journalctl -u vllm.service --since "10 min ago" | grep -iE "error|traceback" | grep -v "no entries" || true
+    journalctl -u {{ 'vllm.service' if item.role == 'primary' else 'vllm-' + item.id + '.service' }} --since "10 min ago" | grep -iE "error|traceback" | grep -v "no entries" || true
   args:
     executable: /bin/bash
+  loop: "{{ vllm_enabled_models }}"
+  loop_control:
+    label: "{{ item.id }}"
   register: vllm_journal_errors
   changed_when: false
   become: true
@@ -112,6 +155,9 @@
 - name: Report journalctl scan result
   ansible.builtin.debug:
     msg: >-
-      {{ 'journalctl clean — no error/traceback lines in the last 10 minutes'
+      {{ item.item.id ~ ': journalctl clean — no error/traceback lines in the last 10 minutes'
-      if vllm_journal_errors.stdout | trim | length == 0
+      if item.stdout | trim | length == 0
-      else 'WARNING — journalctl lines matched error/traceback: ' + vllm_journal_errors.stdout }}
+      else item.item.id ~ ' WARNING — journalctl lines matched error/traceback: ' ~ item.stdout }}
+  loop: "{{ vllm_journal_errors.results }}"
+  loop_control:
+    label: "{{ item.item.id }}"

ansible/roles/deploy-vllm/templates/vllm.service.j2

@@ -34,6 +34,16 @@ ExecStart={{ vllm_venv_path }}/bin/python -m vllm.entrypoints.openai.api_server
     --served-model-name {{ item.id }} \
     --host {{ vllm_serve_host }} \
     --port {{ item.port }} \
+{% if item.role == 'embedding' %}
+    --runner pooling \
+    --convert embed \
+{% endif %}
+{% if item.trust_remote_code is defined and item.trust_remote_code %}
+    --trust-remote-code \
+{% endif %}
+{% if item.enforce_eager is defined and item.enforce_eager %}
+    --enforce-eager \
+{% endif %}
 {% if item.quantization is defined and item.quantization != 'none' %}
     --quantization {{ item.quantization }} \
 {% endif %}
@@ -41,7 +51,11 @@ ExecStart={{ vllm_venv_path }}/bin/python -m vllm.entrypoints.openai.api_server
     --max-model-len {{ item.max_model_len }} \
     --dtype {{ vllm_dtype }} \
     --api-key ${VLLM_API_KEY} \
+{% if item.role != 'embedding' %}
     --enable-prefix-caching
+{% else %}
+    --no-enable-prefix-caching
+{% endif %}
 
 Restart={{ vllm_restart_policy }}
 RestartSec=10

cluster/applications/hindsight/externalsecret.yaml

@@ -55,6 +55,15 @@ spec:
         # env-var-named keys injected via envFrom
         HINDSIGHT_API_MCP_AUTH_TOKEN: "{{ .HINDSIGHT_API_MCP_AUTH_TOKEN }}"
         HINDSIGHT_API_LLM_API_KEY: "{{ .HINDSIGHT_API_LLM_API_KEY }}"
+        # Cutover to vLLM (t_e6facb19, 2026-08-31): vLLM enforces its API key
+        # on every request (unlike llama-swap, which accepted any/no key) —
+        # confirmed empirically, a bad/missing key gets a real 401
+        # {"error":"Unauthorized"}. Both the completions endpoint (:8000) and
+        # the embeddings endpoint (:8020) are separate vLLM processes but
+        # share the SAME key (api-key.env is written once, read by both
+        # systemd units per roles/deploy-vllm/tasks/api-key.yml). Reuse
+        # HINDSIGHT_API_LLM_API_KEY's value for the embeddings key too.
+        HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY: "{{ .HINDSIGHT_API_LLM_API_KEY }}"
   data:
     # 1Password item "hindsight", field "postgres-password".
     # letters+digits only / URL-safe: required because the chart interpolates the
@@ -69,11 +78,15 @@ spec:
       remoteRef:
         key: hindsight
         property: HINDSIGHT_API_MCP_AUTH_TOKEN
-    # 1Password item "nous" (vault mk-labs), field "api-key" (Ryan-provisioned,
+    # Cutover to vLLM (t_e6facb19, 2026-08-31): astro-orbiter's llama-swap
-    # decision 4). Replaces the prior "local-placeholder" value — astro-orbiter
+    # (which needed no real auth) is being retired for Hindsight's
-    # did not validate; now wired to the Nous free-tier inference API. Materialized
+    # completions+embeddings roles. vLLM enforces a real API key — read the
-    # Secret key name stays HINDSIGHT_API_LLM_API_KEY (env-var-named, envFrom).
+    # SAME key roles/deploy-vllm's api-key.yml phase writes to
+    # /etc/vllm/api-key.env on astro-orbiter, sourced from 1Password
+    # op://mk-labs/vllm/api-key (item "vllm", field "api-key", vault mk-labs).
+    # Replaces the prior "nous" item's api-key (Nous free-tier fallback,
+    # no longer the active LLM backend once this cutover lands).
     - secretKey: HINDSIGHT_API_LLM_API_KEY
       remoteRef:
-        key: nous
+        key: vllm
         property: api-key

cluster/applications/hindsight/values.yaml

@@ -60,12 +60,36 @@ existingSecret: hindsight-credentials
 # ----------------------------------------------------------------------------
 api:
   env:
-    # Restore (2026-08-29, t_e0e6f7ca): astro-orbiter back online; move LLM back
+    # Cutover to vLLM (t_e6facb19, 2026-08-31): astro-orbiter's llama-swap is
-    # to local Qwen3.8-27B-Q4_K_M on llama-swap. Nous free tier returns 400
+    # being retired for the completions+embeddings roles Hindsight needs.
-    # 'missing tags' on Hindsight structured fact-extraction (retain broken).
+    # vLLM serves Qwen2.5-32B-Instruct-AWQ (:8000, completions) and
-    HINDSIGHT_API_LLM_BASE_URL: "http://astro-orbiter:8001/v1"
+    # 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: "Qwen3.8-27B-Q4_K_M"
+    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
@@ -74,6 +98,9 @@ api:
     # 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