Skip to content

B200/Qwen: source evidence and execution boundary

Checked 2026-10-18. Hardware inventory, format semantics, model identity and runtime compatibility are separate evidence. This course review used CPU checks and source inspection; it did not run a B200, NVL72 or a 72B model.

Model snapshot

The official Qwen2.5-72B-Instruct config is pinned at 495f39366efef23836d0cfae4fbe635880d2be31. The fetched config, index and model-card hashes are in qwen-metadata.json. The index reports 145,412,407,296 tensor bytes across 37 shards. The metadata identifies the Qwen license, not Apache-2.0. The model's 32K default configuration and the card's context-extension recipe must not be conflated.

This is an explicit historical full-attention teaching model. The newer hybrid case in the quantization course has a different vocabulary, hidden size and state architecture; its memory formula and community quantization export cannot simply replace this model ID.

Platform and software sources

Source What was checked Limit
DGX B200 guide eight GPUs, 1440 GB aggregate HBM actual available memory needs device query
GB200 specification two GPUs/Grace Superchip, 372 GB GPU HBM, NVL72 composition a different SKU from DGX B200; aggregate links are not one CUDA device
CUTLASS 4.8 SM100 operand/scale/accumulator contracts no GPU compilation or kernel result here
vLLM serve documentation current CLI concepts moving docs, backed by pinned source check below
Transformer Engine NVFP4 2.20.2 training recipe documentation not proof of a serving checkpoint loader
TensorRT schemes 11.4 graph semantics distinct from TensorRT-LLM backend and JetPack-supported versions

Software source hashes pin vLLM 0.32.0 at 5df63a7203cdb9328d824a27b682e77c7b9928be and CUTLASS 4.8.0 at 098de2a652cf8f00fd70b2df54051c7eccbb855a. Serving flags were found in the selected parser sources; a full runtime parse/startup or B200 dispatch was not executed. Versioned vLLM documentation URLs returned HTTP 403 during the check, so the course links the accessible stable docs plus immutable source.

Locally executed evidence

The CPU budget record uses existing quantization and collective arithmetic: 10 GiB of BF16 KV at 32K; a hypothetical free-memory budget fits one such sequence but not two; TP=72 fails the simple equal-head partition; normalized AllReduce bandwidth is not its algorithm bandwidth. All calculations passed their stated assertions. They are not hardware fit tests.

The linked FP4 lab, attention lab and Gloo collectives lab contain actual framework/test evidence and their limitations. This pass reuses those fixtures instead of inventing another set of model or GPU results.

Validation and source URL checks record the final checks. Remaining device work includes loading the selected checkpoint, output/quality evaluation, dispatch validation, per-rank memory, NCCL/topology testing, latency/goodput and recovery.