Skip to content

v1.2.0 — Training, LoRA/PEFT, DVC, experiment tracking, LangGraph advisor

Latest
Latest

Choose a tag to compare

View all tags
@hamidmatiny hamidmatiny released this 27 Jul 21:48
v1.2.0
8ee19f9

[1.2.0] — 2026-07-27

v1.2.0 close-out: advanced GPU packaging (validate-only), cost-per-token, training backends,
LoRA/PEFT, DVC for deterministic exports, pluggable experiment tracking, and a tool-grounded
LangGraph advisor with non-fabrication CI. Tag: v1.2.0.
Feature commits:

Phase commits (16–22)

Phase Commit Summary
16 96cb7e1 Advanced GPU serving — GPTQ/AWQ/FP8 + TensorRT-LLM templates
17 164ef1a Cost-per-token + GPU utilization tracking
18 179a1cc Training backends + training-job contract
19 21d590b LoRA/PEFT adapter fine-tuning + transparent serving
20 baa2ba4 DVC for deterministic model exports
21 3e81029 Pluggable experiment tracking (MLflow + W&B offline)
22 c61e16c LangGraph advisor — tool-grounded, non-fabrication

Phase 22 — LangGraph advisor (tool-grounded, non-fabrication)

  • advisor/: LangGraph with query_prometheus, read_benchmark_results, query_routing_history, recommend (ADR-014)
  • Synthesis via template + optional pinned reference-tiny-llm; no paid LLM API in CI
  • CI asserts every numeric/backend claim in the answer appears in that run’s tool evidence (extends ADR-007)

Phase 21 — Pluggable experiment tracking (MLflow + W&B offline)

Feature commit 3e81029; artifact-serve fix 62510a7.

  • training/common/tracking.py: ExperimentTracker interface; MlflowTracker + WandbTracker (offline-only); default VULCAN_TRACKER_BACKEND=none (ADR-013)
  • FSDP/DDP + LoRA report existing loss/throughput through the interface (no recomputation)
  • Compose MLflow on host port 9014; CI asserts MLflow API metrics + W&B offline-run-* dir (never wandb.ai)

Phase 20 — DVC data versioning (deterministic model exports)

Feature commit baa2ba4.

  • dvc.yaml wraps existing export_llm.py / export_vision.py; local filesystem remote only in CI (ADR-012)
  • Cross-check: DVC-tracked primary outs' SHA256 must match sha256sums.txt / MANIFEST (does not replace MANIFEST)
  • Scope: deterministic reference exports only — no training/adapter content-hash tracking (extends ADR-009 / ADR-011)
  • CI dvc-repro job: dvc repro + clean dvc status + unchanged dvc.lock

Phase 19 — LoRA / PEFT (adapter fine-tuning + transparent serving)

Feature commit 21d590b.

  • LoraFineTuneSpec / LoraFineTuneResult in contracts/training-job-contract/ (ADR-011; extends ADR-010)
  • training/fsdp-ddp/lora/: CPU PEFT fine-tune on pinned reference-tiny-llm; structural adapter verify (no SHA256 pin — ADR-009)
  • BentoML optionally serves reference-tiny-llm-lora-demo through unchanged /v1/infer (ADR-001)
  • CI: logits delta proof (base vs base+adapter) on a fixed prompt

Phase 18 — Training backends (v1.2.0 track)

Feature commit 179a1cc.

  • contracts/training-job-contract/: TrainingJobSpec / TrainingJobResult OpenAPI + JSON Schema (ADR-010)
  • training/{ray-train,fsdp-ddp,deepspeed}/: CPU-simulated distributed training (gloo, world_size=2; ADR-009)
  • FSDP/DDP SIGTERM → checkpoint → resume test; DeepSpeed ZeRO-1/2 CPU path + GPU runbook
  • Cost-exporter: vulcan_estimated_cost_usd_per_training_step from training results × ADR-008 $/GPU-hour
  • Host ports 9011–9013 (compose profile training); ADR-009 / ADR-010

Phase 17 — Cost-per-token and GPU utilization tracking

Feature commit 164ef1a.

  • observability/gpu-metrics/: real DCGM-exporter Helm values + cluster Prometheus scrape for phase-7 pools; synthetic DCGM-shaped exporter for compose/CI
  • Cost-exporter: vulcan_estimated_cost_usd_per_token from Bedrock pricing-reference.json and ADR-008 $/GPU-hour × benchmark throughput (phase-7 instance types only)
  • Grafana: cost-per-token + LIVE-SYNTHETIC GPU util panels; LIVE vs PLACEHOLDER table updated
  • ADR-008; KNOWN_GAPS #2 closed; CI smoke asserts cost-per-token + synthetic DCGM

Phase 16 — Advanced GPU serving (v1.1.0 track)

Feature commit 96cb7e1.

  • Extended serving/vllm/docs/gpu-mode.md: continuous batching, PagedAttention vs CPU KV, speculative decoding pairs
  • serving/vllm/gpu-variants/{gptq,awq,fp8}/: schema-valid resource manifests (supports_quantization, declared VRAM envelopes)
  • serving/triton/tensorrt-llm/: TensorRT-LLM config.pbtxt template + Dockerfile + runbook
  • ADR-007; adr-gate; CI validate-only (no GPU, no invented numbers)
Assets 2
Loading