@@ -3,6 +3,7 @@ set -euo pipefail
 
 dockerfile="${1:-docker/Dockerfile.ci}"
 workflow="${2:-.github/workflows/cicd-main.yml}"
+fw_final_dockerfile="${3:-docker/Dockerfile.fw_final}"
 
 baseline_arg_line=$(grep -n '^ARG BASELINE_MCORE_REF$' "$dockerfile" | cut -d: -f1)
 baseline_clone_line=$(grep -n 'git clone --filter=blob:none --no-checkout' "$dockerfile" | cut -d: -f1)
@@ -24,6 +25,15 @@ trap 'rm -rf "$temporary_dir"' EXIT
 ((baseline_line < dispatched_copy_line))
 ((dispatched_copy_line < delta_line))
 
+fw_final_from_line=$(grep -n '^FROM ${NEMO_FW_FINAL_BASE_IMAGE} AS nemo_fw_final$' "$fw_final_dockerfile" | cut -d: -f1)
+fw_final_root_line=$(grep -n '^USER root$' "$fw_final_dockerfile" | cut -d: -f1)
+fw_final_clone_line=$(grep -n '^RUN git clone --depth 1 https://github.com/NVIDIA-NeMo/NeMo.git' "$fw_final_dockerfile" | cut -d: -f1)
+[[ -n "$fw_final_from_line" ]]
+[[ -n "$fw_final_root_line" ]]
+[[ -n "$fw_final_clone_line" ]]
+((fw_final_from_line < fw_final_root_line))
+((fw_final_root_line < fw_final_clone_line))
+
 grep -q -- '--mount=type=cache,target=/root/.cache/uv' "$dockerfile"
 if grep -q -- '--mount=type=secret,id=GH_TOKEN' "$dockerfile"; then
   echo "Baseline MCore clone must not require a GitHub token" >&2
 
@@ -15,6 +15,9 @@
 ARG NEMO_FW_FINAL_BASE_IMAGE
 FROM ${NEMO_FW_FINAL_BASE_IMAGE} AS nemo_fw_final
 
+# The CI base image runs as UID 1000, while framework assembly writes to root-owned system paths.
+USER root
+
 WORKDIR /opt
 
 ##############################################################################
 
@@ -41,7 +41,7 @@ test = ["coverage>=7.8.1"]
 
 [tool.uv.sources]
 "nemo-export-deploy" = { path = "/opt/Export-Deploy", editable = true }
-"nemo-evaluator" = { path = "/opt/Evaluator", editable = true }
+"nemo-evaluator" = { path = "/opt/Evaluator/packages/nemo-evaluator", editable = true }
 "nemo-run" = { path = "/opt/Run", editable = true }
 "lightning" = { path = "stubs/lightning" }
 
 
@@ -82,6 +82,7 @@ training/moe-optimization.md
 training/peft.md
 training/packed-sequences.md
 training/multi-token-prediction.md
+training/dspark-speculative-decoding.md
 training/callbacks.md
 ```
 
 
@@ -4,6 +4,30 @@
 
 DeepSeek V4 models are supported via the Bridge system with auto-detected configuration and weight mapping.
 
+## Chat SFT preprocessing
+
+`DirectHFSFTDatasetConfig` automatically selects the DeepSeek V4 chat collator
+from the tokenizer's `name_or_path`. Because the official V4 tokenizer does not
+provide a Jinja chat template, each training row must select the rendering mode
+explicitly with either `enable_thinking` or `thinking_mode`:
+
+```json
+{
+  "messages": [
+    {"role": "user", "content": "Solve 1 + 1."},
+    {"role": "assistant", "reasoning_content": "Add the terms.", "content": "2"}
+  ],
+  "enable_thinking": true,
+  "truncate_history_thinking": false
+}
+```
+
+Use `enable_thinking: false` for chat mode. Historical reasoning is truncated by
+default in thinking mode; set `truncate_history_thinking: false` to retain it.
+OpenAI-format tool definitions can be
+provided in the row's top-level `tools` field, and tool calls/results remain in
+the structured `messages` list.
+
 ## Model Architecture Features
 
 - **Hybrid Attention (DSv4HybridSelfAttention)**: Per-layer mix of dense MLA and Compressed Sparse Attention selected by `compress_ratios`
 
@@ -189,6 +189,13 @@ dataset.preprocessing = PromptCompletionSFTPreprocessingConfig(
 
 Structured multi-turn rows require chat preprocessing; Bridge does not silently flatten them into prompt-completion text.
 
+Model-specific Jinja controls can be supplied per row under `chat_template_kwargs`, for example
+`{"truncate_history_thinking": false}` to retain all historical reasoning. Megatron Bridge translates this canonical
+control to model-template-specific parameter names where necessary. The data pipeline owns tokenization,
+truncation, padding, and template selection, so those controls cannot be overridden from a dataset row. Tool
+schemas remain in the top-level `tools` field. Processor-batched VLM rows must use identical template kwargs and
+tools within a microbatch.
+
 To train on several sources at once, pass a list to `source`. Each entry in `source_weights` is an epoch count for its source: `1.0` keeps every row once, `2.5` keeps every row twice plus half of them again, and `0.5` keeps half of them. Fractional passes draw without replacement and round up to a whole row. Omitting the weights gives every source one pass.
 
 ```python
 
@@ -114,6 +114,8 @@ See the [peft.sh](peft.sh) script for LoRA fine-tuning with sequence-packing.
 For hosted or local Hugging Face data and a complete one-GPU Qwen3-VL run, start with the
 [Hugging Face multimodal tutorial](../../../../tutorials/data/hf-multimodal/README.md). For sharded WebDataset data,
 use the [multimodal Energon tutorial](../../../../tutorials/data/energon/README.md).
+For a pinned, self-contained real-data example, follow the
+[Nemotron Image Training v3 Energon tutorial](../../../../tutorials/data/energon/nemotron-image-v3.md).
 For a worked image-caption training example, see the
 [DataComp Energon tutorial](../../../../tutorials/data/datacomp/README.md).
 
 
@@ -116,6 +116,10 @@ async def _generate(
     ) as llm:
         if llm.is_primary_rank:
             results = await asyncio.gather(*(llm.generate(prompt, sampling_params) for prompt in prompts))
+            failed_results = [result for result in results if result.failed()]
+            if failed_results:
+                details = ", ".join(f"request {result.request_id}={result.status.name}" for result in failed_results)
+                raise RuntimeError(f"Async inference failed: {details}")
             print_rank_0("======== ASYNC GENERATED TEXT OUTPUT ========")
             for idx, result in enumerate(results):
                 print_rank_0(f"[{idx}] Prompt: {prompts[idx]}")
 
@@ -171,7 +171,9 @@ repository's `ChatMLWebdataset` tar-member contract.
 The [DataComp tutorial](../../tutorials/data/datacomp/README.md) documents a
 complete image-caption preparation and training flow; the
 [multimodal Energon tutorial](../../tutorials/data/energon/README.md) documents
-the general tar-member contract.
+the general tar-member contract. The
+[Nemotron Image Training v3 tutorial](../../tutorials/data/energon/nemotron-image-v3.md) shows a pinned small-subset
+download, conversion, and Qwen3-VL `train.sh` launch through this selector.
 
 ### OpenMathInstruct-2