[fix] support general logic to bypass fp32 downcast and fix qwen35 A_log dtype#975
Merged
[fix] support general logic to bypass fp32 downcast and fix qwen35 A_log dtype#975
Conversation
yueming-yuan
approved these changes
Apr 14, 2026
b96a400 to
87930ef
Compare
Rewrite fp32_param_preservation.md as a practical how-to guide for supporting fp32 parameters in bf16 models. Leads with a 2-step quick start, uses Qwen3.5 A_log as a complete example, and keeps the cast chain analysis as background. Register the doc in index.rst. Made-with: Cursor
GuanxingLu
pushed a commit
to GuanxingLu/miles
that referenced
this pull request
Apr 21, 2026
…log dtype (radixark#975) Co-authored-by: yueming-yuan <yym022502@gmail.com>
DavidBellamy
added a commit
to LLM360/miles
that referenced
this pull request
Apr 21, 2026
…region clusters (#10) * Revert "[BUGFIX] [P2PRDMA] Add rollout post-processing after P2PRDMA weight updates" (radixark#882) * [Fix] fix ci (radixark#894) * Avoid threading for ray getting object (radixark#886) * Add explicit errors for unsupported Megatron profiles (radixark#887) * Add nvfp4 quantizer files (radixark#907) * Bump flash-linear-attention version to 0.4.2 (radixark#892) * [BUGFIX] Invoke "post_process_quantization" by default after weight updating (radixark#890) Co-authored-by: Yueming Yuan <yym022502@gmail.com> * Add heartbeat and id to session server (radixark#866) * fix: adding thin glm5 image to docker build + latest tag sync (radixark#871) * Add consistent hashing routing policy for rollout (radixark#891) Co-authored-by: Yueming Yuan <yueming@Mac.attlocal.net> * [example] add retool v2 example with multi-turn framework interfaces (radixark#654) Co-authored-by: GuanxingLu <gxlu02@gmail.com> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * Expose rollout-batch-size, n-samples-per-prompt, global-batch-size as CLI args in swe-agent-v2 (radixark#954) Co-authored-by: Shi Dong <shi.dong@radixark.ai> * chore: remove obsolete swe-agent server.py and run-qwen3.sh (radixark#952) Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * Add weight staleness control for fully async rollout (radixark#958) * Fix/pause generation mode (radixark#924) Co-authored-by: Yueming Yuan <yym022502@gmail.com> * [v0.5.10][1] Bump sglang to v0.5.10 (radixark#898) * [v0.5.10][2] Fix apply_chat_template behavior for transformers >=5.0 (radixark#926) Co-authored-by: guapisolo <guapisolo@gmail.com> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * [v0.5.10][3] Fix processor return_tensors duplicate kwarg for transformers >=5.0 (radixark#927) Co-authored-by: guapisolo <guapisolo@gmail.com> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * [v0.5.10][4] Fix _no_split_modules set not subscriptable in transformers >=5.0 (radixark#931) * [v0.5.10][5] Disable piecewise cuda graph to avoid NVLS oom (radixark#935) * [v0.5.10][6][FSDP] fix outdated weight update logic in FSDP (radixark#948) Co-authored-by: guapisolo <guapisolo@gmail.com> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Co-authored-by: maocheng23 <35615230+maocheng23@users.noreply.github.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com> * [v0.5.10][7][FSDP] move FSDP to experimental and disable by default (radixark#961) * Add skiplist and more robust calculation on val (radixark#965) * [fix] tiny fix debug rollout only in weight version check (radixark#967) * feat: real cp support with relayout fix for qwen3.5 train/rollout mismatch (radixark#885) * [AMD] Upgrade to sglv0.5.10 (radixark#973) * switch model to actor (radixark#756) * [fix] support general logic to bypass fp32 downcast and fix qwen35 A_log dtype (radixark#975) Co-authored-by: yueming-yuan <yym022502@gmail.com> * fix: populate prefix_cache_info in OpenAI/session rollout path (radixark#960) * Remove prepare_harbor_tasks.py; use harbor-private adapters (radixark#982) * [fix] Skip flush_cache in in_place mode and add fully async example (radixark#974) Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * GLM47 full cmd for async and sync reasoning (radixark#986) Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: handle non-tool appended messages in TITO incremental tokenization (radixark#949) Co-authored-by: Yanbin Jiang <jybsuper@gmail.com> * [docker] Add sgl-model-gateway install and download .tar.gz assets (radixark#895) * [ci] fix hf rate limit error by caching tokenizer loading (radixark#1014) Co-authored-by: maocheng23 <35615230+maocheng23@users.noreply.github.com> * Use load_generate_function in legacy sglang_rollout path (radixark#1016) * Update CODEOWNERS to add new reviewers (radixark#1021) * Support moe lora for gpt-oss (radixark#798) Co-authored-by: Ethan (Yusheng) Su <yushengsu.thu@gmail.com> * [fix] restore expert_bias to fp32 before bridge weight export (radixark#811) * [chore] drop legacy transformers upgrade pin for glm47-flash and qwen35 (radixark#1018) * [fix] Enforce param dtype before wrap ddp (radixark#992) Co-authored-by: Zhichen Zeng <zczeng@uw.edu> * [upgrade] update Megatron-Bridge source and LoRA CI to megatron e2e tests and (radixark#1023) * [CI] Drop --use-miles-router from R3 tests and add r3 comparasion test between sgl & miles router (radixark#1015) * wandb: raise init_timeout, add retry wrapper, fix shared-mode init for cross-region clusters In online + shared mode, both `init_wandb_primary` and `init_wandb_secondary` make HTTPS round-trips to wandb cloud (login + run create/attach). On high-latency cross-region clusters (e.g. Abu Dhabi MBZUAI ↔ wandb-cloud US-West) with concurrent actor bursts, a single round-trip can exceed the wandb SDK's 90s default `init_timeout` — tearing down the whole run with a silent handshake abort. Observed on RL360 job 1564420, which forced `WANDB_MODE=offline` as a global default ever since (see https://github.com/LLM360/RL360/issues/87). The issue's original diagnosis assumed a local primary↔secondary socket handshake race. That's not how shared mode works — per wandb's own feature PR (wandb/wandb#6882), each writer spawns an independent wandb-core that talks to the cloud directly; aggregation is server-side by run_id. No local socket exists. The failure mode is pure network/latency, not a local readiness race. Changes ------- - Bump `init_timeout` to 300s for primary and secondary Settings. Configurable via `WANDB_INIT_TIMEOUT_SECS` env var for tuning. - Wrap both init paths in a bounded exponential-backoff retry (`_wandb_init_with_retry`) that re-attempts on wandb.errors.CommError and wandb.errors.UsageError. 3 attempts with 5→10→20s backoff by default, tunable via `WANDB_INIT_RETRY_ATTEMPTS` / `WANDB_INIT_RETRY_BACKOFF_SECS`. - Add `x_label` tagging per wandb distributed-training docs: primary gets `rank_<rank>_primary`, secondaries get `rank_<rank>_secondary`. Enables per-rank console-log filtering in the wandb UI. - Drop `reinit=True` from secondary init_kwargs. Shared mode natively supports concurrent writers on a single run; `reinit=True` triggered stale-state warnings on secondary actors without functional benefit. Followups this change enables ----------------------------- - `WANDB_MODE=offline` can be removed from scale.yaml's extra_env default once a pilot run confirms online mode boots cleanly. - The tmux-based `~/bin/wandb-sync-rl360.sh` workaround on David's M2 account becomes obsolete (no more offline-only default). - Near-realtime wandb dashboards replace the ~2-minute-lag offline sync; per-rank system metrics via x_label filtering. --------- Co-authored-by: JD <jaedon.guo@gmail.com> Co-authored-by: Ethan (Yusheng) Su <yushengsu.thu@gmail.com> Co-authored-by: fzyzcjy <5236035+fzyzcjy@users.noreply.github.com> Co-authored-by: Ziang Li <ziangli@umich.edu> Co-authored-by: Zhichen Zeng <zczeng@uw.edu> Co-authored-by: JensenFire <xinji1@microsoft.com> Co-authored-by: Yueming Yuan <yym022502@gmail.com> Co-authored-by: maocheng23 <35615230+maocheng23@users.noreply.github.com> Co-authored-by: Douglas Yang <douglasyang88@gmail.com> Co-authored-by: Yueming Yuan <yueming@Mac.attlocal.net> Co-authored-by: Huapeng Zhou <73010314+PopSoda2002@users.noreply.github.com> Co-authored-by: GuanxingLu <gxlu02@gmail.com> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Co-authored-by: Shi-Dong <Shi-Dong@users.noreply.github.com> Co-authored-by: Shi Dong <shi.dong@radixark.ai> Co-authored-by: Jiajun Li <48857426+guapisolo@users.noreply.github.com> Co-authored-by: guapisolo <guapisolo@gmail.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com> Co-authored-by: Yuzhen Zhou <82826991+zyzshishui@users.noreply.github.com> Co-authored-by: Yanbin Jiang <jybsuper@gmail.com> Co-authored-by: Ying Sheng <sqy1415@gmail.com> Co-authored-by: Yisheng Gong <yishenggong9437@gmail.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
PR description generated by cc and reviewed by me.
radixark/Megatron-LM#23 also needed.
FP32 parameter preservation across Megatron's
Float16ModulewrapMotivation
Qwen3.5's
A_logmust stay fp32 through hf → mcore conversion. If it gets rounded to bf16, Megatron-side no longer match sglang's fp32A_logon the rollout side — precision drifts and the train/rollout pair is no longer equivalent. The previouspatch_weight_to_mcore_format_preserve_fp32monkey-patch looked correct but silently shipped bf16.The two cast points
megatron/core/transformer/module.py:440—Float16Modulector runsmodule.bfloat16()nn.Parameter.data→ bf16 at wrap timembridge/core/bridge.py:842—w.to(self.dtype)in_weight_to_mcore_formatself.dtype(bf16)mbridge/core/bridge.py:246—t.to(param.device, dtype=param.dtype)inload_weightsparam.dtype(bf16, due to cast 1)Cast 1 has no declarative opt-out in
nn.Moduleor Megatron — even Megatron's own_maintain_float32_expert_bias(moe/router.py:209-218) uses a post-hoc.data.to(float32)workaround.Why the old patch was insufficient
patch_weight_to_mcore_format_preserve_fp32setself.dtype = Noneduring_weight_to_mcore_format, which handled cast 2 only. It had no handle on cast 1 (Megatron-side) and missed cast 3 entirely — cast 3 readsparam.dtype, which is bf16 because cast 1 already ran. End-to-end,A_logstill shipped as bf16.The replacement — two pieces, each closing one cast point
Downstream (
miles/backends/megatron_utils/fp32_param_utils.py):mark_param_dtype(param, dtype)tags intent at the model definition site;enforce_marked_param_dtypes(model)is called right afterget_model(in bothmiles/backends/megatron_utils/model.py:129andtools/convert_hf_to_torch_dist.py:113) and re-casts tagged params'.databack to fp32, preservingParameteridentity so the optimizer and DDP bucketing registered afterwards see stable tensors.This closes cast 1 and cast 3 in one step — once
param.dtype == fp32, thebridge.py:246in-place cast is a no-op for tagged params.Upstream (
miles_plugins/mbridge/qwen3_5.py:254-261): subclass override inQwen3_5Bridge._weight_to_mcore_formatearly-returnshf_weights[0].to(torch.float32).contiguous()for any name ending inself_attention.linear_attn.A_log(matches MTP layers too). This closes cast 2.Why both pieces are required
From
tools/debug_a_log_old_flow.py:param.dtypecastThe enforce-only row is the trap: dtype is fp32 but the bits were already rounded at cast 2 and up-cast back into an fp32 container. Only a value-level comparison against the HF source catches it — see
test_old_patch_only_regresses_without_enforcein the test file.Extending to other params
mark_param_dtype(self.X, torch.float32)right afterself.X = nn.Parameter(...).self.dtypepre-cast (i.e. the HF ckpt ships it fp32 and you can't afford rounding at cast 2), add a name-matched early-return in the relevant Bridge subclass's_weight_to_mcore_format.No changes to
enforce_marked_param_dtypes, Megatron, or mbridge base.Tests
tests/fast/backends/megatron_utils/test_fp32_param_utils.py— 13 CPU-only tests covering the downstream helper, the upstream bridge override, and a bit-exact end-to-end round-trip plus the negative regression guard against the old patch's failure mode.