Day 0 GB300 DeepSeek-V4-Pro FP4 vLLM disagg #1150

Open

Oseltamivir wants to merge 31 commits intomainfrom

dsv4-fp4-gb300-dynamo-vllm-disagg

Commits on Apr 25, 2026

Port DeepSeek-V4-Pro FP4 disaggregated vLLM sweep from gb200 to gb300-cr
Oseltamivir
committed
Fill in PR link for gb300-cr changelog entry
Oseltamivir
committed
Rename gb300-cr to gb300-cw; fix model path to /mnt/vast/models/dsv4/
Oseltamivir
committed
Fix gb300-cw SLURM account and extend runner group to _2/_3
Oseltamivir
committed
Pin runner-side uv to /tmp so x86 binary doesn't leak to ARM64 compute
Oseltamivir
committed
Force --segment per recipe via sbatch_directives
Oseltamivir
committed
Cap cargo parallelism via CARGO_BUILD_JOBS=4 in gb300 recipes
Oseltamivir
committed
Force --mem=0 (use full node memory) on every gb300 recipe; fix heredoc backtick bug
Oseltamivir
committed
Merge branch 'main' into dsv4-fp4-gb300-dynamo-vllm-disagg
Oseltamivir
authored
Update perf-changelog.yaml
Oseltamivir
authored
Update gb300 recipe headers — segment is recipe-driven, not auto
Oseltamivir
committed
Set NVIDIA_VISIBLE_DEVICES + DRIVER_CAPABILITIES so enroot mounts libcuda.so.1
Oseltamivir
committed
Cache dynamo wheel build globally on /mnt/vast (gb300-cw)
Oseltamivir
committed
Switch dynamo cache lock from flock to mkdir (NFS doesn't honor flock)
Oseltamivir
committed
Pre-build dynamo wheel via single-node srun before sbatch (gb300-cw)
Oseltamivir
committed
Prebuild srun: add --mem=0, cap CARGO_BUILD_JOBS=8, drop rustc debuginfo
Oseltamivir
committed
Mount /mnt/vast/dynamo_cache into worker containers (extra_mount)
Oseltamivir
committed
Patch vllm HANDSHAKE_TIMEOUT_MINS 5->30 in setup script
Oseltamivir
committed
Drop NVL-only NCCL flags + add NCCL_DEBUG=INFO
Oseltamivir
committed
Re-add NCCL_MNNVL_ENABLE, add debug diagnostics, reduce to 1p1d reproducer
Oseltamivir
committed
Remove vLLM HANDSHAKE_TIMEOUT_MINS sed patch from setup script
Oseltamivir
committed
Restore handshake timeout patch, add DP Coordinator logging, drop NCCL_DEBUG_FILE
Oseltamivir
committed
Rewrite coordinator patch to match actual vLLM source strings
Oseltamivir
committed
Rewrite coordinator patch: regex matching + inspect.getsource verify
Oseltamivir
committed
Merge branch 'main' into dsv4-fp4-gb300-dynamo-vllm-disagg
Oseltamivir
authored

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Day 0 GB300 DeepSeek-V4-Pro FP4 vLLM disagg #1150

Day 0 GB300 DeepSeek-V4-Pro FP4 vLLM disagg #1150
Oseltamivir wants to merge 31 commits intomainfrom
dsv4-fp4-gb300-dynamo-vllm-disagg

Commits on Apr 25, 2026

Port DeepSeek-V4-Pro FP4 disaggregated vLLM sweep from gb200 to gb300-cr

Fill in PR link for gb300-cr changelog entry

Rename gb300-cr to gb300-cw; fix model path to /mnt/vast/models/dsv4/

Fix gb300-cw SLURM account and extend runner group to _2/_3

Pin runner-side uv to /tmp so x86 binary doesn't leak to ARM64 compute

Force --segment per recipe via sbatch_directives

Cap cargo parallelism via CARGO_BUILD_JOBS=4 in gb300 recipes

Force --mem=0 (use full node memory) on every gb300 recipe; fix heredoc backtick bug

Merge branch 'main' into dsv4-fp4-gb300-dynamo-vllm-disagg

Update perf-changelog.yaml

Update gb300 recipe headers — segment is recipe-driven, not auto

Set NVIDIA_VISIBLE_DEVICES + DRIVER_CAPABILITIES so enroot mounts libcuda.so.1

Cache dynamo wheel build globally on /mnt/vast (gb300-cw)

Switch dynamo cache lock from flock to mkdir (NFS doesn't honor flock)

Pre-build dynamo wheel via single-node srun before sbatch (gb300-cw)

Prebuild srun: add --mem=0, cap CARGO_BUILD_JOBS=8, drop rustc debuginfo

Mount /mnt/vast/dynamo_cache into worker containers (extra_mount)

Patch vllm HANDSHAKE_TIMEOUT_MINS 5->30 in setup script

Drop NVL-only NCCL flags + add NCCL_DEBUG=INFO

Re-add NCCL_MNNVL_ENABLE, add debug diagnostics, reduce to 1p1d reproducer

Remove vLLM HANDSHAKE_TIMEOUT_MINS sed patch from setup script

Restore handshake timeout patch, add DP Coordinator logging, drop NCCL_DEBUG_FILE

Rewrite coordinator patch to match actual vLLM source strings

Rewrite coordinator patch: regex matching + inspect.getsource verify

Merge branch 'main' into dsv4-fp4-gb300-dynamo-vllm-disagg

Commits on Apr 26, 2026

more

Merge branch 'main' into dsv4-fp4-gb300-dynamo-vllm-disagg

Merge branch 'main' into dsv4-fp4-gb300-dynamo-vllm-disagg

configs

Commits on Apr 28, 2026

PR84 copy

Merge branch 'main' into dsv4-fp4-gb300-dynamo-vllm-disagg