VR200 cfgs, Lm3 70b, 405b, qwen3 30b, 235b, gpt-oss, kimi#3374
Open
malay-nagda wants to merge 4 commits intomainfrom
Open
VR200 cfgs, Lm3 70b, 405b, qwen3 30b, 235b, gpt-oss, kimi#3374malay-nagda wants to merge 4 commits intomainfrom
malay-nagda wants to merge 4 commits intomainfrom
Conversation
Signed-off-by: Malay Nagda <malayn@nvidia.com>
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
Signed-off-by: Malay Nagda <malayn@nvidia.com>
malay-nagda
changed the title
Signed-off-by: Malay Nagda <malayn@nvidia.com>
Contributor
📝 WalkthroughWalkthroughThe PR adds VR200 hardware target support for pretraining configurations across four model families: GPT-OSS 120B, Kimi K2, Llama 70B/405B, and Qwen3 30B/235B. For each model, new VR200-specific configuration factory functions are introduced alongside workload base configuration aliases for multiple precision formats. Changes
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~25 minutes Possibly related PRs
Suggested labels
Suggested reviewers
🚥 Pre-merge checks | ✅ 4✅ Passed checks (4 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Actionable comments posted: 4
🧹 Nitpick comments (1)
scripts/performance/configs/kimi/kimi_llm_pretrain.py (1)
140-177: Consider extracting a shared Kimi pretrain-config builder.This new VR200 function duplicates the same setup path used by other GPU-specific factories. A small internal helper would reduce drift risk between targets.
♻️ Refactor sketch
+def _kimi_k2_pretrain_config_for_gpu( + *, + gpu: str, + precision: str = "bf16", + config_variant: str = "v1", + optimizer_type: str = "muon", +) -> ConfigContainer: + base_cfg = get_workload_base_config( + model_family_name="kimi", + model_recipe_name="kimi_k2", + gpu=gpu, + compute_dtype=precision.upper(), + task="pretrain", + config_variant=config_variant, + ) + cfg = pretrain_config(optimizer_type=optimizer_type) + cfg.mixed_precision = get_precision_config(precision) + if base_cfg.moe_flex_dispatcher_backend is not None: + cfg.model.moe_flex_dispatcher_backend = base_cfg.moe_flex_dispatcher_backend + apply_flex_dispatcher_backend(cfg.model, cfg.model.moe_flex_dispatcher_backend) + if base_cfg.pp_layout: + cfg.model.pipeline_model_parallel_layout = base_cfg.pp_layout + else: + cfg.model.pipeline_model_parallel_layout = _get_kimi_k2_pipeline_layout( + base_cfg.pipeline_model_parallel_size, + base_cfg.virtual_pipeline_model_parallel_size, + ) + set_kimi_k2_common_configs(cfg) + set_workload_base_configs(cfg, base_cfg) + cfg.comm_overlap.overlap_grad_reduce = True + return cfg🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@scripts/performance/configs/kimi/kimi_llm_pretrain.py` around lines 140 - 177, The kimi_k2_pretrain_config_vr200 function duplicates GPU-specific setup logic found in other factory functions; extract a shared builder (e.g., build_kimi_k2_pretrain_config) that accepts GPU-specific params (gpu name, base_cfg) and performs the common steps: call get_workload_base_config, create cfg via pretrain_config, attach mixed precision via get_precision_config, conditionally set cfg.model.moe_flex_dispatcher_backend and call apply_flex_dispatcher_backend, compute or assign pipeline layout (using _get_kimi_k2_pipeline_layout when base_cfg.pp_layout is empty), call set_kimi_k2_common_configs and set_workload_base_configs, and set cfg.comm_overlap.overlap_grad_reduce; then refactor kimi_k2_pretrain_config_vr200 to call this shared builder with vr200-specific args so code paths in functions like kimi_k2_pretrain_config_vr200, apply_flex_dispatcher_backend, set_kimi_k2_common_configs, and set_workload_base_configs remain consistent.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@scripts/performance/configs/gpt_oss/gpt_oss_llm_pretrain.py`:
- Around line 89-113: The function gpt_oss_120b_pretrain_config_vr200 currently
defaults config_variant="v1" which fails VR200 workload lookup; update the
function signature in gpt_oss_120b_pretrain_config_vr200 to use
config_variant="v2" (or remove the default and require the caller to pass the
explicit variant), and ensure any internal uses of config_variant (calls to
get_workload_base_config) continue to pass the corrected value so VR200 entries
in gpt_oss_workload_base_configs.py resolve correctly.
In `@scripts/performance/configs/llama/__init__.py`:
- Around line 36-43: The imported symbols (llama31_405b_pretrain_config_b200,
llama31_405b_pretrain_config_b300, llama31_405b_pretrain_config_gb200,
llama31_405b_pretrain_config_gb300, llama31_405b_pretrain_config_h100,
llama31_405b_pretrain_config_vr200) are flagged as unused (F401); to fix,
explicitly export them by adding an __all__ list that includes each of those
names in the module that currently imports them, or alternatively reference them
in a re-exporting statement so ruff/flake8 recognizes they are intentionally
exposed; update the module's top-level exports accordingly and run ruff
check/format to ensure the lint error is resolved.
In `@scripts/performance/configs/llama/llama3_llm_pretrain.py`:
- Around line 120-153: The function llama3_70b_pretrain_config_vr200 currently
defaults config_variant="v1", which conflicts with VR200 V2-only presets; remove
the arbitrary default so callers must pass an explicit config_variant (change
the signature to config_variant: str with no default), update the function
docstring to mention config_variant is required, and keep the rest of the logic
(get_workload_base_config(..., config_variant=config_variant)) unchanged so
callers must opt into "v2" or other variants explicitly.
In `@scripts/performance/configs/llama/llama31_llm_pretrain.py`:
- Around line 115-148: The function llama31_405b_pretrain_config_vr200 currently
defaults config_variant to "v1" which is incompatible with VR200 presets; change
the config_variant default to "v2" (or remove the default and require callers to
pass the variant) and ensure the get_workload_base_config(...) call uses that
corrected value so workload-base lookup succeeds; optionally add a simple
validation in llama31_405b_pretrain_config_vr200 to raise a clear error if an
unsupported variant (e.g., "v1") is passed.
---
Nitpick comments:
In `@scripts/performance/configs/kimi/kimi_llm_pretrain.py`:
- Around line 140-177: The kimi_k2_pretrain_config_vr200 function duplicates
GPU-specific setup logic found in other factory functions; extract a shared
builder (e.g., build_kimi_k2_pretrain_config) that accepts GPU-specific params
(gpu name, base_cfg) and performs the common steps: call
get_workload_base_config, create cfg via pretrain_config, attach mixed precision
via get_precision_config, conditionally set
cfg.model.moe_flex_dispatcher_backend and call apply_flex_dispatcher_backend,
compute or assign pipeline layout (using _get_kimi_k2_pipeline_layout when
base_cfg.pp_layout is empty), call set_kimi_k2_common_configs and
set_workload_base_configs, and set cfg.comm_overlap.overlap_grad_reduce; then
refactor kimi_k2_pretrain_config_vr200 to call this shared builder with
vr200-specific args so code paths in functions like
kimi_k2_pretrain_config_vr200, apply_flex_dispatcher_backend,
set_kimi_k2_common_configs, and set_workload_base_configs remain consistent.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro Plus
Run ID: b92e10ce-07f1-4bfa-b898-72b3c33f8bce
📒 Files selected for processing (14)
scripts/performance/configs/gpt_oss/__init__.pyscripts/performance/configs/gpt_oss/gpt_oss_llm_pretrain.pyscripts/performance/configs/gpt_oss/gpt_oss_workload_base_configs.pyscripts/performance/configs/kimi/__init__.pyscripts/performance/configs/kimi/kimi_llm_pretrain.pyscripts/performance/configs/kimi/kimi_workload_base_configs.pyscripts/performance/configs/llama/__init__.pyscripts/performance/configs/llama/llama31_llm_pretrain.pyscripts/performance/configs/llama/llama31_workload_base_configs.pyscripts/performance/configs/llama/llama3_llm_pretrain.pyscripts/performance/configs/llama/llama3_workload_base_configs.pyscripts/performance/configs/qwen/__init__.pyscripts/performance/configs/qwen/qwen3_llm_pretrain.pyscripts/performance/configs/qwen/qwen3_workload_base_configs.py
Comment on lines
+89
to
+113
| def gpt_oss_120b_pretrain_config_vr200( | ||
| precision: str = "bf16", mock: bool = True, config_variant: str = "v1" | ||
| ) -> ConfigContainer: | ||
| """VR200, baseline config.""" | ||
| base_cfg = get_workload_base_config( | ||
| model_family_name="gpt_oss", | ||
| model_recipe_name="gpt_oss_120b", | ||
| gpu="vr200", | ||
| compute_dtype=precision.upper(), | ||
| task="pretrain", | ||
| config_variant=config_variant, | ||
| ) | ||
| precision_config = get_precision_config(precision) | ||
|
|
||
| cfg = gpt_oss_120b_pretrain_config() | ||
| cfg.mixed_precision = precision_config | ||
| if base_cfg.moe_flex_dispatcher_backend is not None: | ||
| apply_flex_dispatcher_backend(cfg.model, base_cfg.moe_flex_dispatcher_backend) | ||
| cfg.comm_overlap = CommOverlapConfig(tp_comm_overlap=bool(base_cfg.tensor_model_parallel_size > 1)) | ||
| cfg.comm_overlap.tp_comm_overlap = False if precision == "nvfp4" else cfg.comm_overlap.tp_comm_overlap | ||
| set_gpt_oss_common_configs(cfg) | ||
| set_workload_base_configs(cfg, base_cfg) | ||
|
|
||
| return cfg | ||
|
|
Contributor
There was a problem hiding this comment.
Default config_variant="v1" breaks VR200 lookup.
At Line 90, the default variant is "v1", but VR200 workload entries in scripts/performance/configs/gpt_oss/gpt_oss_workload_base_configs.py are V2-only. A default call will fail config resolution.
💡 Proposed fix
def gpt_oss_120b_pretrain_config_vr200(
- precision: str = "bf16", mock: bool = True, config_variant: str = "v1"
+ precision: str = "bf16", mock: bool = True, config_variant: str = "v2"
) -> ConfigContainer:As per coding guidelines, "Do not add arbitrary defaults for configs; be as explicit as possible."
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@scripts/performance/configs/gpt_oss/gpt_oss_llm_pretrain.py` around lines 89
- 113, The function gpt_oss_120b_pretrain_config_vr200 currently defaults
config_variant="v1" which fails VR200 workload lookup; update the function
signature in gpt_oss_120b_pretrain_config_vr200 to use config_variant="v2" (or
remove the default and require the caller to pass the explicit variant), and
ensure any internal uses of config_variant (calls to get_workload_base_config)
continue to pass the corrected value so VR200 entries in
gpt_oss_workload_base_configs.py resolve correctly.
Comment on lines
36
to
43
| from .llama31_llm_pretrain import ( | ||
| llama31_405b_pretrain_config_b200, | ||
| llama31_405b_pretrain_config_b300, | ||
| llama31_405b_pretrain_config_gb200, | ||
| llama31_405b_pretrain_config_gb300, | ||
| llama31_405b_pretrain_config_h100, | ||
| llama31_405b_pretrain_config_vr200, | ||
| ) |
Contributor
There was a problem hiding this comment.
Fix CI-blocking F401 on llama31_llm_pretrain imports.
These imports are being flagged unused by Flake8 in the current re-export pattern, which will fail lint in CI.
🛠️ Minimal lint fix
- from .llama31_llm_pretrain import (
+ from .llama31_llm_pretrain import ( # noqa: F401
llama31_405b_pretrain_config_b200,
llama31_405b_pretrain_config_b300,
llama31_405b_pretrain_config_gb200,
llama31_405b_pretrain_config_gb300,
llama31_405b_pretrain_config_h100,
llama31_405b_pretrain_config_vr200,
)🧰 Tools
🪛 Flake8 (7.3.0)
[error] 36-36: '.llama31_llm_pretrain.llama31_405b_pretrain_config_b200' imported but unused
(F401)
[error] 36-36: '.llama31_llm_pretrain.llama31_405b_pretrain_config_b300' imported but unused
(F401)
[error] 36-36: '.llama31_llm_pretrain.llama31_405b_pretrain_config_gb200' imported but unused
(F401)
[error] 36-36: '.llama31_llm_pretrain.llama31_405b_pretrain_config_gb300' imported but unused
(F401)
[error] 36-36: '.llama31_llm_pretrain.llama31_405b_pretrain_config_h100' imported but unused
(F401)
[error] 36-36: '.llama31_llm_pretrain.llama31_405b_pretrain_config_vr200' imported but unused
(F401)
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@scripts/performance/configs/llama/__init__.py` around lines 36 - 43, The
imported symbols (llama31_405b_pretrain_config_b200,
llama31_405b_pretrain_config_b300, llama31_405b_pretrain_config_gb200,
llama31_405b_pretrain_config_gb300, llama31_405b_pretrain_config_h100,
llama31_405b_pretrain_config_vr200) are flagged as unused (F401); to fix,
explicitly export them by adding an __all__ list that includes each of those
names in the module that currently imports them, or alternatively reference them
in a re-exporting statement so ruff/flake8 recognizes they are intentionally
exposed; update the module's top-level exports accordingly and run ruff
check/format to ensure the lint error is resolved.
Comment on lines
+120
to
+153
| def llama3_70b_pretrain_config_vr200( | ||
| precision: str = "bf16", mock: bool = True, config_variant: str = "v1" | ||
| ) -> ConfigContainer: | ||
| """VR200, baseline config.""" | ||
| base_cfg = get_workload_base_config( | ||
| model_family_name="llama", | ||
| model_recipe_name="llama3_70b", | ||
| gpu="vr200", | ||
| compute_dtype=precision.upper(), | ||
| task="pretrain", | ||
| config_variant=config_variant, | ||
| ) | ||
| precision_config = get_precision_config(precision) | ||
|
|
||
| if precision == "bf16": | ||
| comm_overlap_cfg = userbuffers_bf16_b200_h8192_tp2_mbs1_seqlen8192 | ||
| else: | ||
| comm_overlap_cfg = userbuffers_fp8_b200_h8192_tp2_mbs1_seqlen8192 | ||
|
|
||
| cfg = llama3_70b_pretrain_config() | ||
| cfg.mixed_precision = precision_config | ||
| set_llama3_common_configs(cfg) | ||
| set_workload_base_configs(cfg, base_cfg) | ||
|
|
||
| if cfg.ddp.use_megatron_fsdp: | ||
| cfg.ddp.fsdp_double_buffer = True | ||
| cfg.model.gradient_accumulation_fusion = False # Disabled to avoid functional errors | ||
| cfg.ddp.suggested_communication_unit_size = 800000000 | ||
|
|
||
| cfg.comm_overlap.tp_comm_overlap_cfg = comm_overlap_cfg | ||
| cfg.comm_overlap.tp_comm_overlap = False if precision == "nvfp4" else cfg.comm_overlap.tp_comm_overlap | ||
|
|
||
| return cfg | ||
|
|
Contributor
There was a problem hiding this comment.
Default VR200 config variant should not be v1.
At Line 121, defaulting config_variant to "v1" conflicts with the newly added VR200 workload presets (V2-only), so default calls will fail.
💡 Proposed fix
def llama3_70b_pretrain_config_vr200(
- precision: str = "bf16", mock: bool = True, config_variant: str = "v1"
+ precision: str = "bf16", mock: bool = True, config_variant: str = "v2"
) -> ConfigContainer:As per coding guidelines, "Do not add arbitrary defaults for configs; be as explicit as possible."
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@scripts/performance/configs/llama/llama3_llm_pretrain.py` around lines 120 -
153, The function llama3_70b_pretrain_config_vr200 currently defaults
config_variant="v1", which conflicts with VR200 V2-only presets; remove the
arbitrary default so callers must pass an explicit config_variant (change the
signature to config_variant: str with no default), update the function docstring
to mention config_variant is required, and keep the rest of the logic
(get_workload_base_config(..., config_variant=config_variant)) unchanged so
callers must opt into "v2" or other variants explicitly.
Comment on lines
+115
to
+148
| def llama31_405b_pretrain_config_vr200( | ||
| precision: str = "bf16", mock: bool = True, config_variant: str = "v1" | ||
| ) -> ConfigContainer: | ||
| """VR200, baseline config.""" | ||
| base_cfg = get_workload_base_config( | ||
| model_family_name="llama", | ||
| model_recipe_name="llama31_405b", | ||
| gpu="vr200", | ||
| compute_dtype=precision.upper(), | ||
| task="pretrain", | ||
| config_variant=config_variant, | ||
| ) | ||
| precision_config = get_precision_config(precision) | ||
|
|
||
| if precision == "bf16": | ||
| comm_overlap_cfg = userbuffers_bf16_b200_h16384_tp4_cp2_mbs1_seqlen8192 | ||
| else: | ||
| comm_overlap_cfg = userbuffers_fp8_b200_h16384_tp4_cp2_mbs1_seqlen8192 | ||
|
|
||
| cfg = llama31_405b_pretrain_config() | ||
| cfg.mixed_precision = precision_config | ||
| set_llama31_common_configs(cfg) | ||
| set_workload_base_configs(cfg, base_cfg) | ||
|
|
||
| if cfg.ddp.use_megatron_fsdp: | ||
| cfg.ddp.fsdp_double_buffer = True | ||
| cfg.model.gradient_accumulation_fusion = False # Disabled to avoid functional errors | ||
| cfg.ddp.num_distributed_optimizer_instances = 2 | ||
|
|
||
| cfg.comm_overlap.tp_comm_overlap_cfg = comm_overlap_cfg | ||
| cfg.comm_overlap.tp_comm_overlap = False if precision == "nvfp4" else cfg.comm_overlap.tp_comm_overlap | ||
|
|
||
| return cfg | ||
|
|
Contributor
There was a problem hiding this comment.
VR200 function default variant is incompatible with available presets.
At Line 116, config_variant defaults to "v1", but VR200 presets for this model are introduced as V2-only. Default calls will fail workload-base lookup.
💡 Proposed fix
def llama31_405b_pretrain_config_vr200(
- precision: str = "bf16", mock: bool = True, config_variant: str = "v1"
+ precision: str = "bf16", mock: bool = True, config_variant: str = "v2"
) -> ConfigContainer:As per coding guidelines, "Do not add arbitrary defaults for configs; be as explicit as possible."
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@scripts/performance/configs/llama/llama31_llm_pretrain.py` around lines 115 -
148, The function llama31_405b_pretrain_config_vr200 currently defaults
config_variant to "v1" which is incompatible with VR200 presets; change the
config_variant default to "v2" (or remove the default and require callers to
pass the variant) and ensure the get_workload_base_config(...) call uses that
corrected value so workload-base lookup succeeds; optionally add a simple
validation in llama31_405b_pretrain_config_vr200 to raise a clear error if an
unsupported variant (e.g., "v1") is passed.
Contributor
|
/claude review |
Contributor
|
LGTM |
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.
What does this PR do ?
Adds support for VR200 system for the models- Llama3 70B, Llama3 405B, Qwen3 30B_a3B, Qwen3 235B_a22B, GPT-OSS 120B and Kimi-K2.
Changelog
GitHub Actions CI
See the CI sectionin the Contributing doc for how to trigger the CI. A Nvidia developer will need to approve and trigger the CI for external contributors.
Before your PR is "Ready for review"
Pre checks:
If you haven't finished some of the above items you can still open "Draft" PR.
Additional Information
Summary by CodeRabbit
New Features