mirror of
https://github.com/deepseek-ai/DeepInfer.git
synced 2026-10-31 08:46:21 +08:00
bf903a91b5
The dependency notes in section 4.4 were a list of things to do by hand; they now name the script that does them, along with what a from-scratch run measured. `h20_e2e_report.sh` sets the two settings a single-node H20 run always needs -- no GIN because the container has no /dev/infiniband, and eager because the Engram gather cannot be captured -- and appends Triton's own cudart to the loader path when it is present, since cu13 images otherwise fail to load any kernel. Requiring a caller to remember three exports to reproduce a run is how a run stops being reproducible.
68 lines
2.7 KiB
Bash
Executable File
68 lines
2.7 KiB
Bash
Executable File
#!/bin/bash
|
|
# Run one end-to-end pass and report only what the pass is judged on.
|
|
#
|
|
# A successful run is not just "some text came out": it is every data-parallel rank agreeing
|
|
# on the token ids, and the fallback list matching what we already know is unported. Both are
|
|
# easy to skim past in the raw log, which prints one block per rank, so they are extracted here.
|
|
#
|
|
# bash scripts/h20_e2e_report.sh
|
|
# EP=8 MAX_TOKENS=64 bash scripts/h20_e2e_report.sh
|
|
#
|
|
# Env in: EP (default 8), MAX_TOKENS (default 32), LOG (default e2e.log),
|
|
# DEEPINFER_HF_MODEL_DIR (required -- the weights).
|
|
#
|
|
# Two settings are defaulted here because single-node H20 always needs them on this image, and
|
|
# both are explained in H20-TRITON-PLAN.md:
|
|
# EP_DISABLE_GIN=1 no /dev/infiniband in the container; SM90 single node uses
|
|
# scale-up, so DeepEP should not look for GIN.
|
|
# DEEPINFER_ENFORCE_EAGER=1 the Engram local gather does a D2H inside CUDA graph capture.
|
|
|
|
set -uo pipefail
|
|
|
|
EP="${EP:-8}"
|
|
MAX_TOKENS="${MAX_TOKENS:-32}"
|
|
LOG="${LOG:-e2e.log}"
|
|
|
|
export DEEPINFER_OPS_BACKEND="${DEEPINFER_OPS_BACKEND:-triton}"
|
|
export EP_DISABLE_GIN="${EP_DISABLE_GIN:-1}"
|
|
export DEEPINFER_ENFORCE_EAGER="${DEEPINFER_ENFORCE_EAGER:-1}"
|
|
|
|
# cu13 images: Triton downloads its own cudart but does not add it to the search path, so any
|
|
# kernel load fails with "libcudart.so.12: cannot open shared object file" without this.
|
|
for d in /root/.triton/nvidia/cudart/*/lib; do
|
|
[ -d "$d" ] && export LD_LIBRARY_PATH="${LD_LIBRARY_PATH:-}:$d"
|
|
done
|
|
|
|
if [ -z "${DEEPINFER_HF_MODEL_DIR:-}" ]; then
|
|
echo "DEEPINFER_HF_MODEL_DIR is not set; point it at the weights directory" >&2
|
|
exit 2
|
|
fi
|
|
|
|
python scripts/h20_ask.py --ep "$EP" --max_tokens "$MAX_TOKENS" >"$LOG" 2>&1
|
|
echo "exit=$?"
|
|
|
|
echo
|
|
echo "=== token ids: distinct sequence and how many ranks produced it"
|
|
# `uniq -c` is the point: a sequence appearing once per rank means every rank agreed.
|
|
grep -o "=== token ids ([0-9]*): \[[^]]*\]" "$LOG" | sed 's/=== token ids ([0-9]*): //' | sort | uniq -c
|
|
|
|
echo
|
|
echo "=== continuation, one per distinct prompt"
|
|
grep -o "=== continuation: '.*'" "$LOG" | sort -u
|
|
|
|
echo
|
|
echo "=== kernels served by the default backend"
|
|
grep -o "\[ops\] kernel '[a-z_0-9]*' used from '[a-z]*'" "$LOG" | sed "s/.*kernel '//;s/' used from.*//" | sort -u
|
|
grep -o "backend '[a-z]*' lacks [0-9]* kernel(s)" "$LOG" | sort -u
|
|
grep -q "rejection sampling" "$LOG" && echo "fused_sample: rejection-sampling variant (MTP verify)"
|
|
|
|
echo
|
|
echo "=== hard failures"
|
|
# Tracebacks specifically: the stats logger mentions "AttributeError" without the run failing.
|
|
if grep -q "Traceback" "$LOG"; then
|
|
echo "TRACEBACK"
|
|
grep -A 5 "Traceback" "$LOG" | head -20
|
|
else
|
|
echo "none"
|
|
fi
|