[WIP]NPU optimizations for Qwen3.5 9B & 35B by Lw135 · Pull Request #308 · redai-studio/Relax

Lw135 · 2026-09-09T09:25:17Z

Summary

Brings the latest SGLang NPU optimizations for Qwen3.5 9B & 35B into ascend-dev-0908 (!70 merge), plus multi-node FLA training support on top: a new Ray head/worker NPU entrypoint, AscendC GDN backend enabled for the 35B-A3B colocate run, and a Dockerfile upgrade to Ascend910_9392 with flash-linear-attention-npu v26.6.0.

Key Changes

1. SGLang NPU Optimization Pack (`sglang-npu.patch` ~3.1k / `sgl-kernel-npu.patch` ~2.9k lines)

GDN / hybrid linear attention NPU backend: recurrent AscendC fusion, QKVZBA packing, exp-race fused Triton kernel, full-attention & TP fusion
MoE: front-routing fusion, weight prefetch, DeepEP token dispatcher updates
Mamba: radix cache extra_buffer scheduling, bf16 SSM dtype
New switches (multi-stream, async exponential, prefill delayer, task queue), fastokens tokenizer, EAGLE speculative-sampling kernels
sgl-kernel-npu: fused fused_qkvzba_conv1d, fused_sigmoid_gating_recurrent, chain/tree speculative kernels; per-target builds for Ascend910_9382

2. Multi-Node FLA Training

New scripts/entrypoint/local-npu-multinode.sh (~145 lines): Ray head/worker cluster bootstrap — head starts Ray and waits for all workers; workers poll the head GCS port and join with retries; delegates to ray-job.sh when RAY_ADDRESS is already set
35B-A3B colocate script: switches to the multinode entrypoint and branches head/worker on MASTER_ADDR == POD_NAME; enables AscendC GDN backend (MINDSPEED_BRIDGE_GDN_BACKEND=ascendc); MODEL_DIR/DATA_DIR default to EXP_DIR; disables --sglang-mamba-scheduler-strategy extra_buffer

3. NPU Dockerfile Upgrades

SOC_VERSION_FOR_SGL_BUILD bumped Ascend910_9382 → Ascend910_9392; patches applied via git am
flash-linear-attention-npu v26.1.0 → v26.6.0; fla_npu now built via pip wheel flow instead of build.sh
Install fastokens 0.3.1

4. Training Scripts

New run-qwen35-9B-8xnpu-colocate.sh: Qwen3.5-9B DAPO-math colocate training (TP4, THD, dynamic batch size, fastokens tokenizer)
Reworked run-qwen35-35B-A3B-16xnpu-colocate.sh: 2×16 NPUs, TP8/PP2/EP16, THD with 20480 tokens/GPU, mem fraction 0.85, --overlap-grad-reduce / --overlap-param-gather, full NPU optimization env flags (SGLANG_NPU_*, ASCEND_USE_FIA, etc.)
loss:

performance:

Env	Job	perf/train_wait_time (s)	perf/train_time (s)	perf/step_time (s)
910C-hdk26.0.rc1	qwen35-35B-A3B(16x910C gbs128) step 1~35 AVG	192.14	81.72	273.86
H800	qwen35-35B-A3B(16xH800 gbs128) step 1~35 AVG	141.71	88.01	229.71
compare	-	0.74	1.08	0.84

5. Runtime Fix

reloadable_process_group.py: skip post-destroy delay on NPU (post_destroy_delay = 0.0)

Major Commits

Commit	Description
`597b703`	!70 merge: add latest sglang optimize & clean code for sglang
`062af1b`	feat(npu): support multi-node FLA training

Merge branch 'wqw_base_lw_dev_0808' into ascend-dev-0808 feat(docker): upgrade NPU env to CANN 9.0.0 with FLA and MindSpeed-Ops # ⭐ Feature ## Upgrade CANN base image and toolchain - Upgrade CANN base image from 8.5.1-a3 to 9.0.0-a3 - Upgrade torch_npu from v2.9.0-7.3.0 to v26.0.1-pytorch2.9.0 - Upgrade triton-ascend from 3.2.0 to 3.2.1 ## Add AscendC Flash Linear Attention (FLA) support - Clone and build fla_npu from flash-linear-attention-npu v26.1.0 - Compile causal_conv1d and gated_delta_rule ops for ascend910_93 ## Add MindSpeed-Ops support - Clone, checkout and install MindSpeed-Ops - Add mindspeed-ops.patch for gated_delta_rule autotune key fix --- # ♻️ Refactor ## Restructure Dockerfile build order - Move torch/torch_npu install before repo clone - Delay COPY . /root/Relax to just before patching - Consolidate Megatron-Bridge into Megatron-LM via cp instead of separate path - Install MindSpeed-Bridge with --no-deps to avoid circular dependency ## Migrate mindspeed-bridge to FLA ops - Replace causal_conv1d from causal_conv1d with FLA implementation - Replace mindspeed_ops l2norm with naive_l2norm fallback - Switch to flash_gated_delta_rule when FLA is available --- # 🐛 Bug Fix ## Fix various compatibility issues - Fix autotune key in gated_delta_rule kernel by removing USE_G and IS_VARLEN - Fix lambda closure bug in transformer_config_init_subclass (use default arg binding) - Fix attention_mask dtype from int32 to bool in model preprocess - Fix vision model config with MoE permute fusion disabled - Handle OmegaConf DictConfig/ListConfig in remove_non_pickleables - Guard apex MixedFusedLayerNorm import with is_npu_available check ## Update sgl-kernel-npu - Upgrade sgl-kernel-npu checkout from 2026.04.15.rc3 to 2026.7.2 - Remove obsolete cherry-pick workaround feat(docker): upgrade NPU env to CANN 9.0.0 with FLA and MindSpeed-Ops feat(docker): upgrade NPU env to CANN 9.0.0 with FLA and MindSpeed-Ops # ⭐ Feature ## Upgrade CANN base image and toolchain - Upgrade CANN base image from 8.5.1-a3 to 9.0.0-a3 - Upgrade torch_npu from v2.9.0-7.3.0 to v26.0.1-pytorch2.9.0 - Upgrade triton-ascend from 3.2.0 to 3.2.1 ## Add AscendC Flash Linear Attention (FLA) support - Clone and build fla_npu from flash-linear-attention-npu v26.1.0 - Compile causal_conv1d and gated_delta_rule ops for ascend910_93 ## Add MindSpeed-Ops support - Clone, checkout and install MindSpeed-Ops - Add mindspeed-ops.patch for gated_delta_rule autotune key fix --- # ♻️ Refactor ## Restructure Dockerfile build order - Move torch/torch_npu install before repo clone - Delay COPY . /root/Relax to just before patching - Consolidate Megatron-Bridge into Megatron-LM via cp instead of separate path - Install MindSpeed-Bridge with --no-deps to avoid circular dependency ## Migrate mindspeed-bridge to FLA ops - Replace causal_conv1d from causal_conv1d with FLA implementation - Replace mindspeed_ops l2norm with naive_l2norm fallback - Switch to flash_gated_delta_rule when FLA is available --- # 🐛 Bug Fix ## Fix various compatibility issues - Fix autotune key in gated_delta_rule kernel by removing USE_G and IS_VARLEN - Fix lambda closure bug in transformer_config_init_subclass (use default arg binding) - Fix attention_mask dtype from int32 to bool in model preprocess - Fix vision model config with MoE permute fusion disabled - Handle OmegaConf DictConfig/ListConfig in remove_non_pickleables - Guard apex MixedFusedLayerNorm import with is_npu_available check ## Update sgl-kernel-npu - Upgrade sgl-kernel-npu checkout from 2026.04.15.rc3 to 2026.7.2 - Remove obsolete cherry-pick workaround [NPU] update sglang fix(megatron): attach _hf_config for MTP bridge when pp>1 # 🐛 Bug Fix ## Fix bridge mapping registry failure with PP > 1 - Set model_bridge._hf_config from bridge.hf_pretrained.config when missing chore(docker): refresh NPU patch set # 🔩 Chore ## Sync NPU patches - Add sgl-kernel-npu.patch (1847 lines) - Update sglang-npu.patch (+1070 lines) - Update mindspeed.patch / mindspeed-bridge.patch / megatron patches build(docker): upgrade NPU image to CANN 9.0.0 # 🔩 Chore ## Upgrade NPU Dockerfile dependencies - Bump base image CANN 8.5.1 → 9.0.0 - Upgrade torch_npu to v26.0.1 (PyTorch 2.9.0) - Upgrade triton-ascend 3.2.0 → 3.2.1 feat(npu): add Qwen3.5 MTP SFT/training scripts feat(npu): add Qwen3.5 MTP training scripts for 0821 # ⭐ Feature ## Add Qwen3.5 MTP training launch scripts - run_qwen35-35B-pokemon-sft-mtp-8xnpu.sh: Qwen3.5-35B-A3B MTP SFT on pokemon-gpt4o-captions, 8xNPU single-node, ray-submit launch - run_qwen35-35B-A3B-16xnpu-colocate-thd.sh: Qwen3.5-35B-A3B colocate THD training on 16xNPU - run_qwen35_9B_mtp_8xnpu_thd.sh: Qwen3.5-9B MTP THD training on 8xNPU --- # 🐛 Bug Fix ## Fix EXP_DIR silently overridden by MODEL_DIR default - EXP_DIR now uses its own default with MODEL_DIR following EXP_DIR, matching the 9B THD script pattern --- # 🔩 Chore ## Align script naming and comments - Rename scripts to the 8xnpu naming convention and drop duplicated .sh suffix - Remove commented-out --qkv-format bshd / --micro-batch-size 1 lines - Sync Usage comments with actual script names Revert "fix(megatron): attach _hf_config for MTP bridge when pp>1" This reverts commit 57809b4. fix: update qwen35 mtp mapping in mindspeed-bridge.patch for relax bridge_converter.py fix(npu): rename Qwen3.5-9B CP script to 16xnpu feat(npu): add Qwen3.5-9B CP training script # ⭐ Feature ## Add Qwen3.5-9B CP colocate training script - Add run-qwen35-9B-8xnpu-cp.sh for DAPO math training with TP4/CP4 on 16 NPUs - Set MINDSPEED_BRIDGE_GDN_USE_TORCH_CONV=1 to use torch conv fallback for GDN ## Support GDN causal conv on NPU via fla_npu - Replace causal_conv1d with fla_npu in mindspeed-bridge.patch GDN layer - Add MINDSPEED_BRIDGE_GDN_USE_TORCH_CONV env switch to fall back to torch conv --- # 🐛 Bug Fix ## Fix repatch ordering before Megatron init - Move repatch(args) before init(args) in MegatronTrainRayActor so bridge patches apply during initialization fix(npu): rename Qwen3.5-9B CP script to 16xnpu # 🐛 Bug Fix ## Fix Qwen3.5-9B CP script naming for 16 NPUs - Rename run-qwen35-9B-8xnpu-cp.sh to run-qwen35-9B-16xnpu-cp.sh to match the actual 16-NPU setup - Update header comment to 16xNPU and log filename to qwen35-9B-GRPO-npu16 docs(npu): add feature support table # 📝 Documentation ## Add feature support table to NPU training guide - Add 特性支持 table covering Qwen3.5-9B CP, Qwen3.5-9B MTP, and Qwen3.5-35B-A3B SFT MTP with reference scripts - Update 下一步 checklist from feature support to performance optimization - Fix MTP row minimum card count to 4卡 to match the script's 8 NPUs fix(gitleaks): drop /mnt/ defaults in NPU scripts # 🔒 Security ## Remove hardcoded internal paths from Qwen3.5 NPU scripts - Replace /mnt/tidalfs-hwwh01 EXP_DIR defaults with the repo-standard `${SCRIPT_DIR}/../../../../exps` in run_qwen35-35B-pokemon-sft-mtp-8xnpu.sh, run_qwen35-35B-A3B-16xnpu-colocate-thd.sh and run_qwen35_9B_mtp_8xnpu_thd.sh - Replace bare /mnt/ placeholder in run-qwen35-9B-16xnpu-cp.sh with the same repo-standard default - Switch the SFT script DATA_DIR default to the `${DATA_DIR:-${EXP_DIR}}` convention used by all other training scripts --- # 🎨 Style ## End-of-file fixer - Remove trailing blank line from docker/Dockerfile.npu Co-Authored-By: Claude <noreply@anthropic.com> update Dockerfile.npu optimizer npu update_weights Add optimized 9B&35B NPU colocate script Add optimized 9B&35B NPU sglang patch Add optimized 9B&35B NPU sgl-kernel-npu patch FIX dockerfile build error Fix sgl-kernel-npu patch [clean code] sglang&kernel patch

feat(NPU): add latest sglang optimize Created-by: llxll Commit-by: liyongwen;jufanbo;wuqiwei;dabuliu123;lixionglong;Tgz27 Merged-by: llxll Description: ## What add latest sglang optimize & clean code for sglang ## Why  ## How  ## Testing  - [ ] `pre-commit run --all-files` passes - [ ] Tests pass (`pytest tests/`) - [ ] New tests added (if applicable) - [ ] Documentation updated (if applicable) ## Type of Change - [ ] Bug fix (non-breaking change that fixes an issue) - [ ] New feature (non-breaking change that adds functionality) - [ ] Breaking change (fix or feature that would cause existing functionality to change) - [ ] Documentation update - [ ] Refactoring (no functional changes) - [ ] Performance improvement - [ ] CI/CD or build changes ## Screenshots / Logs  See merge request: hw-pbclouds/Relax!70

rai-studio-bot

未发现需要阻塞合入的确定性问题；脚本配置问题及修复建议已放在行级评论中。

两份补丁可应用到镜像固定的上游版本，Python 3.10–3.12 CI 测试均通过；pre-commit 仍因两份补丁的末尾换行失败。本次未运行算子数值及多节点训练验证，审查环境缺少所需 NPU 硬件。

^{Powered by Nyanpasu with gpt-6-astra medium, please check the suggestions carefully.}

rai-studio-bot · 2026-09-09T09:39:44Z

-   "${MISC_ARGS[@]}" 2>&1 | tee log/qwen35-35B-MATH-gpu16-sync-${now}.log
+  ray job submit ${RAY_NO_WAIT:+--no-wait} --address="http://${MASTER_ADDR}:8265" \
+  ${WORKING_DIR:+--working-dir "${WORKING_DIR}"} \
+  --runtime-env-json="${RUNTIME_ENV_JSON}" \


优先级：P2

请在提交前将新增优化开关合入任务的 env_vars，两份脚本都需要覆盖。连接已有 Ray 集群时，local-npu.sh 会委托 ray-job-npu.sh；后者的 RUNTIME_ENV_JSON 只包含固定名单，没有这里新增的 SGLANG_NPU_*、SGLANG_GMM2_TRITON 和 MINDSPEED_BRIDGE_GDN_USE_TORCH_CONV 等变量。提交 shell 中的 export 不会更新已运行的 raylet 环境，导致该启动路径下部分优化实际未启用。可在环境初始化完成后、ray job submit 前按以下方式合并：

RUNTIME_ENV_JSON="$(python3 - <<'PY' import json, os env = json.loads(os.environ["RUNTIME_ENV_JSON"]) extra = { "CPU_AFFINITY_CONF", "TORCH_HCCL_ZERO_COPY", "MULTI_STREAM_MEMORY_REUSE", "HCCL_OP_EXPANSION_MODE", "ASCEND_USE_FIA", "GDN_ATTN_BACKEND_TRITON", "STREAMS_PER_DEVICE", "TASK_QUEUE_ENABLE", } env.setdefault("env_vars", {}).update({ k: v for k, v in os.environ.items() if k.startswith(("SGLANG_", "MINDSPEED_BRIDGE_")) or k in extra }) print(json.dumps(env)) PY )"

feat(npu): support multi-node FLA training # ⭐ Feature ## Add multi-node NPU entrypoint - Add `scripts/entrypoint/local-npu-multinode.sh` for Ray head/worker cluster setup - Head node starts Ray head and waits until all worker nodes join - Worker nodes poll head GCS port and join the cluster with retries ## Update Qwen3.5-35B colocate training script - Switch entrypoint to local-npu-multinode.sh and branch head/worker on `MASTER_ADDR` - Enable AscendC GDN backend via `MINDSPEED_BRIDGE_GDN_BACKEND=ascendc` - Default `MODEL_DIR`/`DATA_DIR` to `EXP_DIR` - Disable `--sglang-mamba-scheduler-strategy extra_buffer` ## Upgrade NPU docker image - Bump `SOC_VERSION_FOR_SGL_BUILD` to Ascend910_9392 - Upgrade flash-linear-attention-npu v26.1.0 → v26.6.0 - Replace build.sh flow with pip wheel build flow for fla_npu --- # 🎨 Style ## Normalize patch file endings - Remove trailing blank lines in docker/npu_patch (pre-commit end-of-file-fixer) fix(npu): scope colocate script to master node # 🐛 Bug Fix ## Scope master-node training block - Indent config and ray job submit under the master-node branch - Remove dead worker-node else branch ## Disable SGLang radix cache - Enable --sglang-disable-radix-cache in SGLANG_ARGS --- # 🎨 Style ## Clean up training script - Remove stale packing and micro-batch comments - Normalize trailing whitespace and file endings (pre-commit)

feat(npu): support multi-node FLA training Created-by: dabuliu123 Commit-by: dabuliu123 Merged-by: llxll Description: Multi-node NPU training support. ## Add multi-node NPU entrypoint - Add `scripts/entrypoint/local-npu-multinode.sh` for Ray head/worker cluster setup - Head node starts Ray head and waits until all worker nodes join - Worker nodes poll head GCS port and join the cluster with retries ## Update Qwen3.5-35B colocate training script - Switch entrypoint to local-npu-multinode.sh and branch head/worker on `MASTER_ADDR` - Enable AscendC GDN backend via `MINDSPEED_BRIDGE_GDN_BACKEND=ascendc` - Default `MODEL_DIR`/`DATA_DIR` to `EXP_DIR` - Disable `--sglang-mamba-scheduler-strategy extra_buffer` ## Upgrade NPU docker image - Bump `SOC_VERSION_FOR_SGL_BUILD` to Ascend910_9392 - Upgrade flash-linear-attention-npu v26.1.0 -> v26.6.0 - Replace build.sh flow with pip wheel build flow for fla_npu See merge request: hw-pbclouds/Relax!71

rai-studio-bot

已复查新增的多节点入口和 FLA 构建调整。发现一项会让已有 Ray 集群启动路径静默跳过训练提交的回归，建议修复后再合入；具体证据和可应用修复见新增行级评论。

此前的参数拼写问题已被禁用 radix cache 的新配置取代；环境变量传递和 9B 路径默认值问题仍未解决，沿用原线程。补丁末尾换行问题已修复，当前 CI 全部通过。已完成脚本语法检查及隔离启动模拟；未执行完整 NPU 镜像构建，算子数值和训练集成验证因缺少多节点 NPU 硬件而未运行。

优先级：P3 非行级：PR 描述尚未同步新增提交。描述仍写 Ascend910_9382，并将 35B 的配置概括为 radix cache / extra_buffer，与当前代码不符。请更新为 Ascend910_9392、FLA v26.6.0 wheel 构建及 35B 禁用 radix cache，并补充新多节点入口所需的 MASTER_ADDR、POD_NAME、HOST_IP 配置和对应验证结果，便于复现。

^{Powered by Nyanpasu with gpt-6-astra medium, please check the suggestions carefully.}

rai-studio-bot · 2026-09-10T22:10:41Z

+if [ -z "${RELAX_ENTRYPOINT_MODE:-}" ]; then
+    source "${SCRIPT_DIR}/../../entrypoint/local-npu-multinode.sh"
+fi
+if [ "$MASTER_ADDR" = "$POD_NAME" ]; then


优先级：P1

请保留已有 Ray 集群入口的提交路径。ray-job-npu.sh 会将 MASTER_ADDR 设置为节点 IP，而 POD_NAME 未设置或为 Pod 名称，因此这里的比较会跳过整个提交块。显式通过该入口启动，以及新 helper 检测到已有集群后委托该入口，都会受影响。隔离模拟 RELAX_ENTRYPOINT_MODE=ray-job、MASTER_ADDR=192.0.2.10、未设置 POD_NAME 时，实际脚本退出码为 0，却没有调用 ray job submit。

多节点 helper 的 worker 分支已经阻塞在 sleep inf；返回调用方的已初始化入口应继续提交，无需再次比较 Pod 名称：

Suggested change

if [ "$MASTER_ADDR" = "$POD_NAME" ]; then

if [ -n "${RELAX_ENTRYPOINT_MODE:-}" ]; then

rai-studio-bot

描述已补充多节点入口、FLA v26.6.0 和新构建方式，上一轮描述建议已部分落实。代码未变化，已有集群下跳过训练提交的问题仍在，详见原行级评论。参数拼写问题已被新配置取代；环境变量传递问题仍未解决，9B 路径线程虽已标记解决，当前代码中的默认值仍未补齐。本次没有新增行级评论，CI 全部通过。

优先级：P3 非行级：描述仍有未同步内容。请将第 1 节的内核构建目标 Ascend910_9382 改为 Ascend910_9392，在第 2 节明确 35B 启用了 --sglang-disable-radix-cache，并补充 MASTER_ADDR、POD_NAME、HOST_IP 的配置示例及新多节点路径的验证结果。这样可避免按旧目标构建，也能明确现有性能数据是否覆盖此次启动方式。

^{Powered by Nyanpasu with gpt-6-astra medium, please check the suggestions carefully.}

Michael-Salon · 2026-09-11T08:49:18Z

建议修改：
P1 — 35B 脚本在已有 Ray 集群模式下会跳过训练。
run-qwen35-35B-A3B-16xnpu-colocate.sh (line 67)
新增条件要求 MASTER_ADDR == POD_NAME 才提交任务，但 ray-job-npu.sh 将 MASTER_ADDR 设置为节点 IP，而 POD_NAME 是名称或未设置。通过已有集群入口运行时，条件为假，整个训练提交块被跳过。应仅在自行启动多节点集群时判断节点角色，已有集群模式直接提交。

P1 — 标称 8 卡的 9B 脚本实际申请 16 卡。
run-qwen35-9B-8xnpu-colocate.sh (line 196)
Actor、Rollout 资源和 --num-gpus-per-node 均设置为 16，可见设备也配置为 0–15。在实际 8 卡节点上无法满足资源申请；代码会阻塞在 placement group 的 ready() 等待。应统一改为 8 卡，或明确改成可配置的 16 卡 recipe。

P2 — 9B 脚本缺少模型和数据目录默认值。
run-qwen35-9B-8xnpu-colocate.sh (line 69)
只初始化了 EXP_DIR，随后直接使用 MODEL_DIR、DATA_DIR，被引用的本地入口也未设置它们。未额外导出变量时，路径会变成 /Qwen3.5-9B/ 和 /dapo-math-17k/...，无法读取通常放在实验目录下的文件。建议像其他 recipe 一样默认回退到 EXP_DIR，并检查路径存在。

rai-studio-bot

已复查新提交：此前阻塞合入的 35B 已有集群提交回归已修复，9B 模型和数据目录默认值也已补齐，显式覆盖仍有效。两份脚本语法检查和隔离提交模拟通过，CI 全部通过；缺少多节点 NPU 硬件，未运行训练集成验证。

本轮未发现新的阻塞问题，没有新增行级评论。旧参数拼写问题已被新配置取代，环境变量传递问题仍沿用原线程。

优先级：P3 非行级：描述建议仍部分未落实。请将内核构建目标统一为 Ascend910_9392，明确 35B 禁用 radix cache，并将角色判断说明更新为“由多节点入口负责，已有集群直接提交”；同时补充 MASTER_ADDR、POD_NAME、HOST_IP 配置示例及多节点验证结果，确保说明与修复后的启动流程一致。

^{Powered by Nyanpasu with gpt-6-astra medium, please check the suggestions carefully.}

Lw135 and others added 2 commits September 9, 2026 10:08

Lw135 requested review from NINGBENZHE and Yangruipis as code owners September 9, 2026 09:25

Lw135 changed the title ~~Ascend dev 0908~~ [WIP]SGLang NPU optimizations for Qwen3.5 9B & 35B Sep 9, 2026

rai-studio-bot reviewed Sep 9, 2026

View reviewed changes

Lw135 and others added 2 commits September 10, 2026 16:55

rai-studio-bot suggested changes Sep 10, 2026

View reviewed changes

Lw135 changed the title ~~[WIP]SGLang NPU optimizations for Qwen3.5 9B & 35B~~ [WIP]NPU optimizations for Qwen3.5 9B & 35B Sep 11, 2026

rai-studio-bot reviewed Sep 11, 2026

View reviewed changes

fix npu script

7b79972

rai-studio-bot reviewed Sep 11, 2026

View reviewed changes

Navigation Menu

Uh oh!

[WIP]NPU optimizations for Qwen3.5 9B & 35B - #308

[WIP]NPU optimizations for Qwen3.5 9B & 35B#308
Lw135 wants to merge 5 commits into
redai-studio:mainfrom
Lw135:ascend-dev-0908

Lw135 commented Sep 9, 2026 •

edited

Loading

Uh oh!

rai-studio-bot left a comment

Uh oh!

Uh oh!

rai-studio-bot Sep 9, 2026

Uh oh!

Uh oh!

rai-studio-bot left a comment

Uh oh!

rai-studio-bot Sep 10, 2026

Uh oh!

rai-studio-bot left a comment

Uh oh!

Michael-Salon commented Sep 11, 2026

Uh oh!

rai-studio-bot left a comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

	if [ "$MASTER_ADDR" = "$POD_NAME" ]; then
	if [ -n "${RELAX_ENTRYPOINT_MODE:-}" ]; then

Uh oh!

Conversation

Lw135 commented Sep 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Key Changes

1. SGLang NPU Optimization Pack (sglang-npu.patch ~3.1k / sgl-kernel-npu.patch ~2.9k lines)

2. Multi-Node FLA Training

3. NPU Dockerfile Upgrades

4. Training Scripts

5. Runtime Fix

Major Commits

Uh oh!

rai-studio-bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

rai-studio-bot Sep 9, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

rai-studio-bot left a comment

Choose a reason for hiding this comment

Uh oh!

rai-studio-bot Sep 10, 2026

Choose a reason for hiding this comment

Uh oh!

rai-studio-bot left a comment

Choose a reason for hiding this comment

Uh oh!

Michael-Salon commented Sep 11, 2026

Uh oh!

rai-studio-bot left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Lw135 commented Sep 9, 2026 •

edited

Loading

1. SGLang NPU Optimization Pack (`sglang-npu.patch` ~3.1k / `sgl-kernel-npu.patch` ~2.9k lines)