Skip to content

dsv4-fp4-b300-sglang-mtp: pass --dsv4 to use DSv4 chat template #1182

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Oseltamivir merged 4 commits into from
Apr 27, 2026
Merged

dsv4-fp4-b300-sglang-mtp: pass --dsv4 to use DSv4 chat template #1182

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
ee44ab5
dsv4-fp4-b300-sglang-mtp: pass --dsv4 to use DSv4 chat template
cquil11 Apr 26, 2026
a351d19
perf-changelog: fill in PR #1182 link
cquil11 Apr 26, 2026
5cb6b21
perf-changelog: trim dsv4-mtp entry
cquil11 Apr 26, 2026
74ee570
Merge branch 'main' into chore/sgl-dsv4-b300-mtp-2
Oseltamivir Apr 27, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 8 additions & 1 deletion benchmarks/single_node/dsv4_fp4_b300_sglang_mtp.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -128,6 +128,12 @@ wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$S

pip install -q datasets pandas

# --dsv4 routes prompts through encoding_dsv4.py (PR #1153), which emits the
# <bos><User>...<Assistant><think> framing DeepSeek-V4-Pro expects. The DSv4-Pro
# tokenizer ships without a jinja chat_template, so plain --use-chat-template
# would crash; --dsv4 sidesteps that and satisfies the AGENTS.md rule that all
# MTP scripts must benchmark against chat-formatted inputs (EAGLE acceptance
# silently regresses on raw random tokens).
run_benchmark_serving \
--model "$MODEL" \
--port "$PORT" \
Expand All @@ -138,7 +144,8 @@ run_benchmark_serving \
--num-prompts $((CONC * 10)) \
--max-concurrency "$CONC" \
--result-filename "$RESULT_FILENAME" \
--result-dir "$PWD/"
--result-dir "$PWD/" \
--dsv4

if [ "${RUN_EVAL}" = "true" ]; then
run_eval --framework lm-eval --port "$PORT"
Expand Down
8 changes: 8 additions & 0 deletions perf-changelog.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1903,3 +1903,11 @@
- "ISL=8192: TP4 conc 4-64; DP4 (dp-attn) conc 128-1024; DP8 (dp-attn) conc 1024-8192"
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1155

- config-keys:
- dsv4-fp4-b300-sglang-mtp
description:
- "Pass --dsv4 (PR #1153) to run_benchmark_serving so prompts go through the DeepSeek-V4 chat template (encoding_dsv4.py)"
- "Restores AGENTS.md compliance: every MTP script must benchmark against chat-formatted inputs; PR #1166 had to drop --use-chat-template because the DSv4-Pro tokenizer has no jinja template, which is exactly what --dsv4 sidesteps"
- "EAGLE acceptance rate is sensitive to input distribution; raw random tokens silently regress acceptance vs. chat-framed prompts"
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX

Loading