Skip to content
Original file line number Diff line number Diff line change
Expand Up @@ -33,7 +33,7 @@

model:
path: "deepseek-v4-pro"
container: "lmsysorg/sglang-staging:deepseek-v4-grace-blackwell-dev"
container: "lmsysorg/sglang:nightly-dev-cu13-20260514-f7efff32"

Check warning on line 36 in benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-10p1d-dep4-dep16-14-c8192.yaml

View check run for this annotation

Claude / Claude Code Review

Stale block comment about container alias after container bump

The file-level block comment in all 6 modified GB300 recipe yamls still claims the container was "restored to the aliases mapped in launch_gb300.sh's srtslurm.yaml (`lmsysorg/sglang:deepseek-v4-grace-blackwell` and `deepseek-v4-pro`)", but after this PR the container is the literal pinned tag `lmsysorg/sglang:nightly-dev-cu13-20260514-f7efff32` — no longer the grace-blackwell alias. Nit / documentation-only; consider updating the comment to reflect the new pinning rationale (or just drop the con
Comment thread
Fridge003 marked this conversation as resolved.
Outdated
precision: "fp4"

dynamo:
Expand Down Expand Up @@ -85,6 +85,8 @@
SGLANG_OPT_USE_TOPK_V2: "1"
SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE: "1"
SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
SGLANG_OPT_FIX_HASH_MEGA_MOE: "1"
SGLANG_OPT_USE_FAST_MASK_EP: "1"
SGLANG_OPT_FIX_MEGA_MOE_MEMORY: "1"
Expand Down Expand Up @@ -117,6 +119,8 @@
SGLANG_OPT_USE_JIT_INDEXER_METADATA: "1"
SGLANG_OPT_USE_TOPK_V2: "1"
SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE: "1"
SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
SGLANG_OPT_FIX_HASH_MEGA_MOE: "1"
SGLANG_OPT_USE_FAST_MASK_EP: "1"
SGLANG_OPT_FIX_MEGA_MOE_MEMORY: "1"
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -33,7 +33,7 @@ name: "disagg-gb300-12p1d-dep4-dep12-15-c21504"

model:
path: "deepseek-v4-pro"
container: "lmsysorg/sglang-staging:deepseek-v4-grace-blackwell-dev"
container: "lmsysorg/sglang:nightly-dev-cu13-20260514-f7efff32"
precision: "fp4"

dynamo:
Expand Down Expand Up @@ -85,6 +85,8 @@ backend:
SGLANG_OPT_USE_TOPK_V2: "1"
SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE: "1"
SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
SGLANG_OPT_FIX_HASH_MEGA_MOE: "1"
SGLANG_OPT_USE_FAST_MASK_EP: "1"
SGLANG_OPT_FIX_MEGA_MOE_MEMORY: "1"
Expand Down Expand Up @@ -117,6 +119,8 @@ backend:
SGLANG_OPT_USE_JIT_INDEXER_METADATA: "1"
SGLANG_OPT_USE_TOPK_V2: "1"
SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE: "1"
SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
SGLANG_OPT_FIX_HASH_MEGA_MOE: "1"
SGLANG_OPT_USE_FAST_MASK_EP: "1"
SGLANG_OPT_FIX_MEGA_MOE_MEMORY: "1"
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -33,7 +33,7 @@ name: "disagg-gb300-1p1d-dep4-dep16-5-c1024"

model:
path: "deepseek-v4-pro"
container: "lmsysorg/sglang-staging:deepseek-v4-grace-blackwell-dev"
container: "lmsysorg/sglang:nightly-dev-cu13-20260514-f7efff32"
precision: "fp4"

dynamo:
Expand Down Expand Up @@ -85,6 +85,8 @@ backend:
SGLANG_OPT_USE_TOPK_V2: "1"
SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE: "1"
SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
SGLANG_OPT_FIX_HASH_MEGA_MOE: "1"
SGLANG_OPT_USE_FAST_MASK_EP: "1"
SGLANG_OPT_FIX_MEGA_MOE_MEMORY: "1"
Expand Down Expand Up @@ -117,6 +119,8 @@ backend:
SGLANG_OPT_USE_JIT_INDEXER_METADATA: "1"
SGLANG_OPT_USE_TOPK_V2: "1"
SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE: "1"
SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
SGLANG_OPT_FIX_HASH_MEGA_MOE: "1"
SGLANG_OPT_USE_FAST_MASK_EP: "1"
SGLANG_OPT_FIX_MEGA_MOE_MEMORY: "1"
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -33,7 +33,7 @@ name: "disagg-gb300-1p1d-tp4-tp4-2-c1"

model:
path: "deepseek-v4-pro"
container: "lmsysorg/sglang-staging:deepseek-v4-grace-blackwell-dev"
container: "lmsysorg/sglang:nightly-dev-cu13-20260514-f7efff32"
precision: "fp4"

# See ../1k1k/disagg-gb200-1p1d-dep8-tep8.yaml for the dynamo pin
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -33,7 +33,7 @@ name: "disagg-gb300-4p1d-dep4-dep16-8-c1024"

model:
path: "deepseek-v4-pro"
container: "lmsysorg/sglang-staging:deepseek-v4-grace-blackwell-dev"
container: "lmsysorg/sglang:nightly-dev-cu13-20260514-f7efff32"
precision: "fp4"

dynamo:
Expand Down Expand Up @@ -85,6 +85,8 @@ backend:
SGLANG_OPT_USE_TOPK_V2: "1"
SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE: "1"
SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
SGLANG_OPT_FIX_HASH_MEGA_MOE: "1"
SGLANG_OPT_USE_FAST_MASK_EP: "1"
SGLANG_OPT_FIX_MEGA_MOE_MEMORY: "1"
Expand Down Expand Up @@ -117,6 +119,8 @@ backend:
SGLANG_OPT_USE_JIT_INDEXER_METADATA: "1"
SGLANG_OPT_USE_TOPK_V2: "1"
SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE: "1"
SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
SGLANG_OPT_FIX_HASH_MEGA_MOE: "1"
SGLANG_OPT_USE_FAST_MASK_EP: "1"
SGLANG_OPT_FIX_MEGA_MOE_MEMORY: "1"
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -33,7 +33,7 @@ name: "disagg-gb300-8p1d-dep4-dep16-12-c4096"

model:
path: "deepseek-v4-pro"
container: "lmsysorg/sglang-staging:deepseek-v4-grace-blackwell-dev"
container: "lmsysorg/sglang:nightly-dev-cu13-20260514-f7efff32"
precision: "fp4"

dynamo:
Expand Down Expand Up @@ -85,6 +85,8 @@ backend:
SGLANG_OPT_USE_TOPK_V2: "1"
SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE: "1"
SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
SGLANG_OPT_FIX_HASH_MEGA_MOE: "1"
SGLANG_OPT_USE_FAST_MASK_EP: "1"
SGLANG_OPT_FIX_MEGA_MOE_MEMORY: "1"
Expand Down Expand Up @@ -117,6 +119,8 @@ backend:
SGLANG_OPT_USE_JIT_INDEXER_METADATA: "1"
SGLANG_OPT_USE_TOPK_V2: "1"
SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE: "1"
SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
SGLANG_OPT_FIX_HASH_MEGA_MOE: "1"
SGLANG_OPT_USE_FAST_MASK_EP: "1"
SGLANG_OPT_FIX_MEGA_MOE_MEMORY: "1"
Expand Down
7 changes: 7 additions & 0 deletions perf-changelog.yaml
Original file line number Diff line number Diff line change
Expand Up @@ -2475,4 +2475,11 @@
- "Turn to tp=4 for best perf"
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1375

- config-keys:
- dsv4-fp4-gb300-dynamo-sglang
description:
- "Enable W4A4 (MXFP4) megamoe by appending SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS=1 and SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND=1 wherever SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE is set"
Comment thread
Fridge003 marked this conversation as resolved.
Outdated
- "Update SGLang image from lmsysorg/sglang-staging:deepseek-v4-grace-blackwell-dev to lmsysorg/sglang:nightly-dev-cu13-20260514-f7efff32"
Comment thread
Fridge003 marked this conversation as resolved.
Outdated
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX

Check warning on line 2483 in perf-changelog.yaml

View check run for this annotation

Claude / Claude Code Review

Placeholder pr-link XXXX in perf-changelog.yaml entry

The new perf-changelog.yaml entry at line 2483 still has the placeholder PR link `https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX` — every other entry in this file uses a real PR number, and you flagged this yourself as a TODO in the test plan ("Update placeholder PR link in perf-changelog.yaml"). Please replace `XXXX` with `1382` before merging so the changelog entry resolves to this PR.
Comment thread
Fridge003 marked this conversation as resolved.
Outdated
Comment thread
Fridge003 marked this conversation as resolved.
Outdated