MILAB-6480: optimize embedding resource requests with fluent formulas API

.changeset/sequence-embeddings-resource-formulas.md

@@ -0,0 +1,16 @@
+---
+"@platforma-open/milaboratories.sequence-embeddings.software": minor
+"@platforma-open/milaboratories.sequence-embeddings.workflow": patch
+"@platforma-open/milaboratories.sequence-embeddings.model": patch
+"@platforma-open/milaboratories.sequence-embeddings.ui": patch
+"@platforma-open/milaboratories.sequence-embeddings": minor
+---
+
+Optimize embedding compute-resource requests via the fluent `exec.formula` API (workflow-tengo 6.7.x) and size the GPU batch from the allocated VRAM.
+
+- **Workflow** — the embedding exec now declares resources with `.resources({ onCPU | onGPU })` instead of a flat 16 CPU / 32 GiB / 16 GiB-VRAM request:
+  - GPU path: host CPU/RAM cut to 2 cores / 8 GiB (single streaming process; the accelerator does the work, per-chunk CPU work is interleaved, and the heavy PColumn→TSV shaping is already upstream CPU execs), and VRAM is model-tiered — 6 GiB for the wider checkpoints (ESM-2 650M, CurrAb, PeptideCLM-2), 3 GiB otherwise. This fits the cheapest fractional-L4 tiers (gpu-3g `g6f.xlarge` / gpu-6g `g6f.2xlarge`) instead of forcing the 4×L40S `g6e.12xlarge` the old 16 CPU / 32 GiB / 16 GiB-VRAM request required.
+  - CPU path: cores scale with the batch's token volume (`size("batch")`, 4–16) and RAM tracks the core count (2 GiB/core, 8–32 GiB), with a `.staticFallback` equal to the old fixed 16 / 32 GiB for backends that cannot evaluate resource formulas. Advanced-settings overrides still win per dimension.
+  - The run-report counting pass now sizes its RAM from the source-TSV size (it full-loads the TSV), replacing the flat 4 GiB that could OOM on large inputs.
+- **Software** — on CUDA the per-forward token budget is now auto-sized from the allocated VRAM (`PLATFORMA_GPU_MEMORY`), mirroring how `--max-memory-gb` sizes the host-RAM path: a larger VRAM request yields larger batches (higher throughput), a smaller one stays safe. An explicit `--token-budget` still wins, and the halve-on-OOM retry remains the backstop.
+- **Model / UI** — the `mem` / `cpu` Advanced-Settings fields are now opt-in: a new block leaves them unset so the workflow's automatic sizing applies, and a user value overrides it per dimension. (Existing projects keep any value they already had.)

model/src/dataModel.ts

@@ -53,7 +53,7 @@ export const blockDataModel = new DataModelBuilder()
     // The selection is seeded by the UI on first input connection (specialist-first);
     // starts blank so the dropdowns have an object to bind to before an input exists.
     embedding: {},
-    // Resource defaults for the embedding step (Advanced Settings).
-    mem: 32,
-    cpu: 16,
+    // mem/cpu are intentionally left UNSET: the workflow sizes the embedding step's
+    // resources automatically from device (CPU/GPU) and input volume. They become
+    // opt-in overrides — set only when the user fills them in Advanced Settings.
   }));

model/src/types.ts

@@ -206,8 +206,8 @@ export type BlockDataV1 = {
   scopesInitializedForAnchor?: string;
   /**
    * Advanced — host resources for the embedding step (Resource Allocation in
-   * Advanced Settings). `mem` is GiB, `cpu` is cores. Undefined → workflow
-   * defaults (32 GiB / 16 cores).
+   * Advanced Settings). `mem` is GiB, `cpu` is cores. Undefined → workflow sizes
+   * automatically from device (CPU/GPU) and input volume.
    */
   mem?: number;
   cpu?: number;
@@ -227,7 +227,7 @@ export type BlockDataV2 = {
    * Init-guard: canonical id of the anchor whose default cards were last seeded.
    */
   embeddingsInitializedForAnchor?: string;
-  /** Advanced resource overrides (GiB / cores); undefined → workflow defaults. */
+  /** Advanced resource overrides (GiB / cores); undefined → workflow sizes automatically from device + input. */
   mem?: number;
   cpu?: number;
   defaultBlockLabel?: string;
@@ -249,7 +249,7 @@ export type BlockDataV3 = {
    * input change.
    */
   embeddingInitializedForAnchor?: string;
-  /** Advanced resource overrides (GiB / cores); undefined → workflow defaults. */
+  /** Advanced resource overrides (GiB / cores); undefined → workflow sizes automatically from device + input. */
   mem?: number;
   cpu?: number;
   defaultBlockLabel?: string;
@@ -279,7 +279,7 @@ export type BlockArgs = {
   inputAnchor: PlRef;
   /** Tasks to emit — a 1-element list projected from `data.embedding` (validated complete + compatible). */
   embeddings: EmbeddingTask[];
-  /** Advanced resource overrides for the embedding step (GiB / cores); undefined → workflow defaults. */
+  /** Advanced resource overrides for the embedding step (GiB / cores); undefined → workflow sizes automatically. */
   mem?: number;
   cpu?: number;
 };