[#5860][autodeploy] GPT-OSS MXFP4 support V2: patch before export #7452
Conversation
Signed-off-by: Frida Hou <[email protected]>
📝 WalkthroughWalkthroughAdds a new MXFP4 Triton-backed Torch custom op and exports it via custom_ops. Updates HF model build flow to run a quantization pre-processing pass (dtype update and preprocess) before returning the model. No public API signatures changed beyond the new custom op registration. Changes
Sequence Diagram(s)sequenceDiagram
autonumber
actor U as Caller
participant T as Torch Custom Op<br/>auto_deploy::mxfp4_mlp
participant H as Kernel Hub (_hub)
participant R as Router/Gating
participant S as Swizzle MXFP4
participant M1 as Matmul (gate_up)
participant Act as Fused Activation (swiglu)
participant M2 as Matmul (down)
participant O as Output
U->>T: mxfp4_mlp(hidden_states, weights, scales, biases, top_k, alpha, limit)
T->>H: get_kernel hub (cached)
T->>R: compute router_logits and routing_data
T->>S: swizzle gate_up/down weights + scales (MXFP4 layout)
T->>M1: routed matmul(hidden_states, gate_up) + bias
M1-->>T: intermediate_cache1
T->>Act: apply swiglu(alpha, limit)
Act-->>T: activated_cache
T->>M2: routed matmul(activated_cache, down) + bias
M2-->>T: routed_out
T->>O: reshape to [B,S,H]
O-->>U: tensor
note over T,H: Uses PrecisionConfig / FlexCtx from hub
sequenceDiagram
autonumber
actor C as AutoModelForCausalLMFactory._build_model
participant M as Model
participant Q as AutoHfQuantizer
C->>M: model.eval()
C->>Q: construct from model_config.quantization_config<br/>(pre_quantized=True)
C->>Q: update_dtype(model_config.dtype)
Q-->>C: dtype
C->>M: model.to(updated_dtype)
C->>Q: preprocess_model(model, device_map=None,<br/>keep_in_fp32_modules, config, use_kernels=False)
Q-->>C: preprocessed model
C-->>C: print(model)
C-->>C: return model
Estimated code review effort🎯 4 (Complex) | ⏱️ ~60 minutes Suggested reviewers
✨ Finishing Touches
🧪 Generate unit tests
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. 🪧 TipsChatThere are 3 ways to chat with CodeRabbit:
SupportNeed help? Create a ticket on our support page for assistance with any issues or questions. CodeRabbit Commands (Invoked using PR/Issue comments)Type Other keywords and placeholders
Status, Documentation and Community
|
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 6
🧹 Nitpick comments (6)
tensorrt_llm/_torch/auto_deploy/models/hf.py (1)
1-1: Add 2025 NVIDIA copyright header at file top.Per repo guidelines, prepend the NVIDIA copyright header (current year) above the module docstring.
Apply this diff:
+# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved. + """Interface to initialize and load HF models."""tensorrt_llm/_torch/auto_deploy/custom_ops/mxfp4.py (3)
1-1: Add 2025 NVIDIA copyright header.Apply this diff:
+# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved. + import torch
79-81: Device guard may fail on non-CUDA tensors.with torch.cuda.device(router_logits.device) will error if tensors are on CPU/MPS. Consider using the tensor’s device without entering a CUDA device context.
Apply this diff:
- with torch.cuda.device(router_logits.device): - routing_data, gather_idx, scatter_idx = routing(router_logits, top_k) + routing_data, gather_idx, scatter_idx = routing(router_logits, top_k)
70-76: Validate inputs early (top_k bounds, dimensions).Add checks for top_k in [1, local_experts] and dim compat between hidden_states and router_weight.
Apply this diff:
- hub = _hub() + hub = _hub() @@ - router_logits = torch.nn.functional.linear(hidden_states, router_weight, router_bias) + assert 1 <= top_k <= router_weight.shape[0], "top_k must be within [1, num_experts]" + assert router_weight.shape[1] == hidden_size, "router_weight H mismatch" + router_logits = torch.nn.functional.linear(hidden_states, router_weight, router_bias)Also applies to: 94-96
tensorrt_llm/_torch/auto_deploy/custom_ops/__init__.py (2)
1-1: Add 2025 NVIDIA copyright header.Apply this diff:
+# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved. + """Custom ops and make sure they are all registered."""
9-9: Star re-export is fine here, but consider narrowing public surface later.Importing everything from mxfp4 is consistent with this module, but a curated all can reduce accidental symbol leakage.
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (3)
tensorrt_llm/_torch/auto_deploy/custom_ops/__init__.py(1 hunks)tensorrt_llm/_torch/auto_deploy/custom_ops/mxfp4.py(1 hunks)tensorrt_llm/_torch/auto_deploy/models/hf.py(1 hunks)
🧰 Additional context used
📓 Path-based instructions (4)
**/*
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Filenames compiled into a target must be case-insensitively unique
Files:
tensorrt_llm/_torch/auto_deploy/custom_ops/__init__.pytensorrt_llm/_torch/auto_deploy/models/hf.pytensorrt_llm/_torch/auto_deploy/custom_ops/mxfp4.py
**/*.{h,hpp,hh,hxx,cc,cpp,cxx,cu,cuh,py}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Use spaces, not tabs; indent 4 spaces
Files:
tensorrt_llm/_torch/auto_deploy/custom_ops/__init__.pytensorrt_llm/_torch/auto_deploy/models/hf.pytensorrt_llm/_torch/auto_deploy/custom_ops/mxfp4.py
**/*.py
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
**/*.py: Code must target Python 3.8+
Indent with 4 spaces; do not use tabs (Python)
Maintain module namespace on import: prefer from package.subpackage import foo; use foo.Symbol()
Python filenames use snake_case
Python class names use PascalCase
Python functions and methods use snake_case
Python local variables use snake_case; if starting with a number concept, prefix with k (e.g., k_99th_percentile)
Python global variables use G_ prefix with UPPER_SNAKE_CASE
Python constants use UPPER_SNAKE_CASE
Avoid shadowing variables from outer scopes
Initialize all externally visible class members in init
For public interfaces, prefer docstrings over comments; comments should be for in-function or file-local interfaces
Use Google-style docstrings for classes and functions (Sphinx-parsable)
Document attributes and variables inline with docstrings immediately after assignment
Avoid reflection when a non-reflective approach suffices
Limit except clauses to specific exceptions where possible
When using try/except for duck-typing, keep try body minimal and move logic to else
Files:
tensorrt_llm/_torch/auto_deploy/custom_ops/__init__.pytensorrt_llm/_torch/auto_deploy/models/hf.pytensorrt_llm/_torch/auto_deploy/custom_ops/mxfp4.py
**/*.{cpp,cc,cxx,h,hpp,hh,hxx,cu,cuh,py}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Prepend NVIDIA copyright header (current year) to all source files
Files:
tensorrt_llm/_torch/auto_deploy/custom_ops/__init__.pytensorrt_llm/_torch/auto_deploy/models/hf.pytensorrt_llm/_torch/auto_deploy/custom_ops/mxfp4.py
🧬 Code graph analysis (2)
tensorrt_llm/_torch/auto_deploy/models/hf.py (2)
tensorrt_llm/_torch/auto_deploy/custom_ops/attention_interface.py (1)
to(336-340)tensorrt_llm/_torch/auto_deploy/shim/interface.py (1)
to(33-37)
tensorrt_llm/_torch/auto_deploy/custom_ops/mxfp4.py (1)
cpp/tensorrt_llm/kernels/trtllmGenKernels/blockScaleMoe/runner.h (2)
top_k(221-221)intermediate_size(226-226)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)
- GitHub Check: Pre-commit Check
🔇 Additional comments (2)
tensorrt_llm/_torch/auto_deploy/models/hf.py (1)
186-192: Double-check attribute access to private _keep_in_fp32_modules.This is a private attribute and may not exist across models. The proposed guard above mitigates it; please confirm downstream paths don’t rely on strict type of this value.
tensorrt_llm/_torch/auto_deploy/custom_ops/mxfp4.py (1)
106-117: Ignore typo concerns: API uses gather_indx/scatter_indx and gate_scal
Verified that matmul_ogs() in kernels-community/triton_kernels indeed defines gather_indx and scatter_indx, and the RoutingData field is named gate_scal (no “e”). [1][2]Likely an incorrect or invalid review comment.
| def _hub(): | ||
| global _triton_kernels_hub | ||
| if _triton_kernels_hub is None: | ||
| from kernels import get_kernel | ||
|
|
||
| _triton_kernels_hub = get_kernel("kernels-community/triton_kernels") | ||
| return _triton_kernels_hub | ||
|
|
There was a problem hiding this comment.
🛠️ Refactor suggestion
Harden hub loading with clear error if kernels are missing.
Raise a helpful error when kernels-community/triton_kernels is unavailable.
Apply this diff:
def _hub():
global _triton_kernels_hub
if _triton_kernels_hub is None:
- from kernels import get_kernel
-
- _triton_kernels_hub = get_kernel("kernels-community/triton_kernels")
+ try:
+ from kernels import get_kernel
+ except Exception as e:
+ raise ImportError(
+ "Required Triton kernels hub 'kernels-community/triton_kernels' not found. "
+ "Install or make it discoverable in PYTHONPATH."
+ ) from e
+ _triton_kernels_hub = get_kernel("kernels-community/triton_kernels")
return _triton_kernels_hub📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| def _hub(): | |
| global _triton_kernels_hub | |
| if _triton_kernels_hub is None: | |
| from kernels import get_kernel | |
| _triton_kernels_hub = get_kernel("kernels-community/triton_kernels") | |
| return _triton_kernels_hub | |
| def _hub(): | |
| global _triton_kernels_hub | |
| if _triton_kernels_hub is None: | |
| try: | |
| from kernels import get_kernel | |
| except Exception as e: | |
| raise ImportError( | |
| "Required Triton kernels hub 'kernels-community/triton_kernels' not found. " | |
| "Install or make it discoverable in PYTHONPATH." | |
| ) from e | |
| _triton_kernels_hub = get_kernel("kernels-community/triton_kernels") | |
| return _triton_kernels_hub |
🤖 Prompt for AI Agents
In tensorrt_llm/_torch/auto_deploy/custom_ops/mxfp4.py around lines 6 to 13, the
hub loader blindly imports from kernels and calls get_kernel which can raise
ModuleNotFoundError/ImportError or a lookup failure; wrap the import and
get_kernel call in a try/except that catches ImportError/ModuleNotFoundError and
any exception from get_kernel, then raise a clear RuntimeError (or re-raise
ImportError) with a helpful message that "kernels-community/triton_kernels" is
unavailable and include the original exception text and actionable guidance
(e.g., how to install or configure the kernels package); ensure you still set or
return _triton_kernels_hub when successful.
| """ | ||
| Wrapper that forwards to your Python reference implementation. | ||
| Return: | ||
| routed_out: same leading shape as hidden_states, last dim = H | ||
| router_logits: [T, E] (T = number of tokens = prod(hidden_states.shape[:-1])) | ||
| """ |
There was a problem hiding this comment.
Return docstring is inconsistent with the actual return type.
Docstring advertises router_logits as a return value, but the function only returns routed_out.
Apply this diff to fix the docstring:
- Return:
- routed_out: same leading shape as hidden_states, last dim = H
- router_logits: [T, E] (T = number of tokens = prod(hidden_states.shape[:-1]))
+ Returns:
+ routed_out: same leading shape as hidden_states, last dim = H📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| """ | |
| Wrapper that forwards to your Python reference implementation. | |
| Return: | |
| routed_out: same leading shape as hidden_states, last dim = H | |
| router_logits: [T, E] (T = number of tokens = prod(hidden_states.shape[:-1])) | |
| """ | |
| """ | |
| Wrapper that forwards to your Python reference implementation. | |
| Returns: | |
| routed_out: same leading shape as hidden_states, last dim = H | |
| """ |
🤖 Prompt for AI Agents
In tensorrt_llm/_torch/auto_deploy/custom_ops/mxfp4.py around lines 63 to 68,
the docstring claims the function returns both routed_out and router_logits but
the function only returns routed_out; update the docstring to remove the
router_logits entry and only describe routed_out (its shape: same leading shape
as hidden_states with last dim = H) and the return semantics so the docstring
matches the actual return value.
| batch_size = hidden_states.shape[0] | ||
| intermediate_size = gate_up_blocks.shape[1] // 2 | ||
| hidden_size = hidden_states.shape[-1] | ||
| hidden_states = hidden_states.reshape(-1, hidden_size) | ||
| router_logits = torch.nn.functional.linear(hidden_states, router_weight, router_bias) |
There was a problem hiding this comment.
Preserve input leading shape (2D vs 3D) when reshaping output.
Current code always reshapes to [batch_size, -1, H], adding an extra dim for 2D inputs.
Apply this diff:
- batch_size = hidden_states.shape[0]
- intermediate_size = gate_up_blocks.shape[1] // 2
- hidden_size = hidden_states.shape[-1]
- hidden_states = hidden_states.reshape(-1, hidden_size)
+ orig_shape = tuple(hidden_states.shape[:-1])
+ intermediate_size = gate_up_blocks.shape[1] // 2
+ hidden_size = hidden_states.shape[-1]
+ hidden_states = hidden_states.reshape(-1, hidden_size)
@@
- routed_out = routed_out.reshape(batch_size, -1, hidden_size)
+ routed_out = routed_out.reshape(*orig_shape, hidden_size)Also applies to: 139-140
🤖 Prompt for AI Agents
In tensorrt_llm/_torch/auto_deploy/custom_ops/mxfp4.py around lines 73-77 (and
similarly at 139-140), the code flattens hidden_states to 2D and then
reconstructs outputs but always returns a 3D tensor, which wrongly adds an extra
sequence dimension for original 2D inputs; preserve the input's original ndim by
capturing orig_shape = hidden_states.shape and orig_ndim = hidden_states.dim()
before reshaping, perform the linear op on the flattened tensor, then reshape
router_logits (and any other outputs) back to orig_shape: if orig_ndim == 3 use
(batch_size, -1, feat) otherwise use (batch_size, feat) (or equivalently reshape
to (*orig_shape[:-1], feat)); apply the same pattern for the code at lines
139-140 so 2D inputs remain 2D and 3D inputs remain 3D.
| local_experts = gate_up_blocks.size(0) | ||
| gate_up_blocks = gate_up_blocks.view(local_experts, intermediate_size * 2, -1) | ||
| triton_gate_up_w, gate_up_w_scale_raw = _swizzle_mxfp4( | ||
| gate_up_blocks.transpose(-2, -1), gate_up_scales.transpose(-2, -1) | ||
| ) | ||
| triton_gate_up_w.shape = torch.Size([local_experts, hidden_size, intermediate_size * 2]) |
There was a problem hiding this comment.
🛠️ Refactor suggestion
Make layout-agnostic with explicit dim detection instead of .view().
Inputs may be [E, 2I, H] or [E, H, 2I] (or [E, I, H]/[E, H, I]). Using view assumes a particular memory layout. Use permute based on which dim equals hidden_size.
Apply this diff:
- local_experts = gate_up_blocks.size(0)
- gate_up_blocks = gate_up_blocks.view(local_experts, intermediate_size * 2, -1)
+ local_experts = gate_up_blocks.size(0)
+ # normalize gate_up to [E, 2I, H]
+ if gate_up_blocks.shape[-1] == hidden_size:
+ gu = gate_up_blocks # [E, 2I, H]
+ else:
+ gu = gate_up_blocks.permute(0, 2, 1).contiguous() # [E, 2I, H]
+ gate_up_blocks = gu
@@
- down_blocks = down_blocks.view(local_experts, -1, intermediate_size)
+ # normalize down to [E, I, H]
+ if down_blocks.shape[-1] == hidden_size:
+ dn = down_blocks # [E, I, H]
+ else:
+ dn = down_blocks.permute(0, 2, 1).contiguous() # [E, I, H]
+ down_blocks = dnAlso applies to: 101-105
🤖 Prompt for AI Agents
In tensorrt_llm/_torch/auto_deploy/custom_ops/mxfp4.py around lines 94 to 99
(and similarly lines 101-105), the code uses .view() assuming the tensor layout
[E, 2I, H]; instead detect which dimension equals hidden_size and call .permute
to move dims into [E, 2I, H] ordering before reshaping so the operation is
layout-agnostic. Concretely: check gate_up_blocks.dim() and which
gate_up_blocks.size(dim) == hidden_size, permute to put hidden_size as the last
dim (or the expected position) then reshape to [local_experts, intermediate_size
* 2, -1] (or set shape to [local_experts, hidden_size, intermediate_size * 2]
after swizzle) and apply the same detection+permute fix to the analogous block
at lines 101-105.
| down_blocks = down_blocks.view(local_experts, -1, intermediate_size // 2) | ||
| triton_down_w, down_w_scale_raw = _swizzle_mxfp4( |
There was a problem hiding this comment.
Shape bug: down path uses intermediate_size // 2; should be intermediate_size.
This halves I and will mis-shape weights/biases.
Apply this diff:
- down_blocks = down_blocks.view(local_experts, -1, intermediate_size // 2)
+ down_blocks = down_blocks.view(local_experts, -1, intermediate_size)📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| down_blocks = down_blocks.view(local_experts, -1, intermediate_size // 2) | |
| triton_down_w, down_w_scale_raw = _swizzle_mxfp4( | |
| down_blocks = down_blocks.view(local_experts, -1, intermediate_size) | |
| triton_down_w, down_w_scale_raw = _swizzle_mxfp4( |
🤖 Prompt for AI Agents
In tensorrt_llm/_torch/auto_deploy/custom_ops/mxfp4.py around lines 100-101, the
down path reshapes use intermediate_size // 2 which incorrectly halves the
intermediate dimension causing mis-shaped weights/biases; change the view call
to use intermediate_size (not divided by 2) so down_blocks =
down_blocks.view(local_experts, -1, intermediate_size) and ensure any downstream
code that depends on that shape (e.g., calls to _swizzle_mxfp4 and associated
scale/weight unpacking) expects and processes the full intermediate_size.
| # TODO(fridah): this should be handled by QuantConfigReader | ||
| from transformers.quantizers import AutoHfQuantizer | ||
|
|
||
| hf_quantizer = AutoHfQuantizer.from_config( | ||
| model_config.quantization_config, | ||
| pre_quantized=True, | ||
| ) | ||
|
|
||
| dtype = hf_quantizer.update_dtype(model_config.dtype) | ||
| model.to(dtype) | ||
|
|
||
| hf_quantizer.preprocess_model( | ||
| model=model, | ||
| device_map=None, | ||
| keep_in_fp32_modules=model._keep_in_fp32_modules, | ||
| config=model.config, | ||
| use_kernels=False, | ||
| ) | ||
| print("Model after pre-processing: ", model) | ||
|
|
There was a problem hiding this comment.
🛠️ Refactor suggestion
Guard missing quantization_config, avoid local ImportError, and use logger instead of print.
- If model_config.quantization_config is absent/None, this block will raise.
- Local import may raise ImportError at runtime; surface a clear message.
- model._keep_in_fp32_modules may not exist; guard with getattr.
- Replace print with ad_logger.debug.
Apply this diff:
- # TODO(fridah): this should be handled by QuantConfigReader
- from transformers.quantizers import AutoHfQuantizer
-
- hf_quantizer = AutoHfQuantizer.from_config(
- model_config.quantization_config,
- pre_quantized=True,
- )
-
- dtype = hf_quantizer.update_dtype(model_config.dtype)
- model.to(dtype)
-
- hf_quantizer.preprocess_model(
- model=model,
- device_map=None,
- keep_in_fp32_modules=model._keep_in_fp32_modules,
- config=model.config,
- use_kernels=False,
- )
- print("Model after pre-processing: ", model)
+ # TODO(fridah): route via QuantConfigReader once available.
+ quant_cfg = getattr(model_config, "quantization_config", None)
+ if quant_cfg is None:
+ ad_logger.debug("No quantization_config found; skipping HF quantizer preprocess.")
+ else:
+ try:
+ from transformers.quantizers import AutoHfQuantizer # type: ignore[attr-defined]
+ except Exception as e:
+ raise RuntimeError(
+ "transformers.quantizers.AutoHfQuantizer is required for pre-quantized models."
+ ) from e
+
+ hf_quantizer = AutoHfQuantizer.from_config(quant_cfg, pre_quantized=True)
+ dtype = hf_quantizer.update_dtype(model_config.dtype)
+ model.to(dtype)
+
+ keep_in_fp32_modules = getattr(model, "_keep_in_fp32_modules", ())
+ hf_quantizer.preprocess_model(
+ model=model,
+ device_map=None,
+ keep_in_fp32_modules=keep_in_fp32_modules,
+ config=model.config,
+ use_kernels=False,
+ )
+ ad_logger.debug("Model after quantizer pre-processing: %s", model.__class__.__name__)📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| # TODO(fridah): this should be handled by QuantConfigReader | |
| from transformers.quantizers import AutoHfQuantizer | |
| hf_quantizer = AutoHfQuantizer.from_config( | |
| model_config.quantization_config, | |
| pre_quantized=True, | |
| ) | |
| dtype = hf_quantizer.update_dtype(model_config.dtype) | |
| model.to(dtype) | |
| hf_quantizer.preprocess_model( | |
| model=model, | |
| device_map=None, | |
| keep_in_fp32_modules=model._keep_in_fp32_modules, | |
| config=model.config, | |
| use_kernels=False, | |
| ) | |
| print("Model after pre-processing: ", model) | |
| # TODO(fridah): route via QuantConfigReader once available. | |
| quant_cfg = getattr(model_config, "quantization_config", None) | |
| if quant_cfg is None: | |
| ad_logger.debug("No quantization_config found; skipping HF quantizer preprocess.") | |
| else: | |
| try: | |
| from transformers.quantizers import AutoHfQuantizer # type: ignore[attr-defined] | |
| except Exception as e: | |
| raise RuntimeError( | |
| "transformers.quantizers.AutoHfQuantizer is required for pre-quantized models." | |
| ) from e | |
| hf_quantizer = AutoHfQuantizer.from_config(quant_cfg, pre_quantized=True) | |
| dtype = hf_quantizer.update_dtype(model_config.dtype) | |
| model.to(dtype) | |
| keep_in_fp32_modules = getattr(model, "_keep_in_fp32_modules", ()) | |
| hf_quantizer.preprocess_model( | |
| model=model, | |
| device_map=None, | |
| keep_in_fp32_modules=keep_in_fp32_modules, | |
| config=model.config, | |
| use_kernels=False, | |
| ) | |
| ad_logger.debug("Model after quantizer pre-processing: %s", model.__class__.__name__) |
🤖 Prompt for AI Agents
In tensorrt_llm/_torch/auto_deploy/models/hf.py around lines 175 to 194, the
block assumes model_config.quantization_config exists, does a local import
without handling ImportError, accesses model._keep_in_fp32_modules directly, and
uses print; change it to first check if model_config.quantization_config is
truthy and skip the whole quantization path if not, wrap the local import of
AutoHfQuantizer in a try/except that raises or logs a clear ImportError message,
call AutoHfQuantizer.from_config only when config is present, use getattr(model,
"_keep_in_fp32_modules", None) when passing keep_in_fp32_modules, and replace
the print with ad_logger.debug to log the post-preprocess model state.
|
Will land this feature with #7451 |
1 participant
This PR does the following:
env requirements for huggingface to load MXFP4 model
Summary by CodeRabbit
Description
Test Coverage
PR Checklist
Please review the following before submitting your PR:
PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.
PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.
Test cases are provided for new code paths (see test instructions)
Any new dependencies have been scanned for license and vulnerabilities
CODEOWNERS updated if ownership changes
Documentation updated as needed
The reviewers assigned automatically/manually are appropriate for the PR.
Please check this after reviewing the above items as appropriate for this PR.
GitHub Bot Help
/bot [-h] ['run', 'kill', 'skip', 'reuse-pipeline'] ...Provide a user friendly way for developers to interact with a Jenkins server.
Run
/bot [-h|--help]to print this help message.See details below for each supported subcommand.
Details
run [--reuse-test (optional)pipeline-id --disable-fail-fast --skip-test --stage-list "A10-PyTorch-1, xxx" --gpu-type "A30, H100_PCIe" --test-backend "pytorch, cpp" --add-multi-gpu-test --only-multi-gpu-test --disable-multi-gpu-test --post-merge --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx" --detailed-log --debug(experimental)]Launch build/test pipelines. All previously running jobs will be killed.
--reuse-test (optional)pipeline-id(OPTIONAL) : Allow the new pipeline to reuse build artifacts and skip successful test stages from a specified pipeline or the last pipeline if no pipeline-id is indicated. If the Git commit ID has changed, this option will be always ignored. The DEFAULT behavior of the bot is to reuse build artifacts and successful test results from the last pipeline.--disable-reuse-test(OPTIONAL) : Explicitly prevent the pipeline from reusing build artifacts and skipping successful test stages from a previous pipeline. Ensure that all builds and tests are run regardless of previous successes.--disable-fail-fast(OPTIONAL) : Disable fail fast on build/tests/infra failures.--skip-test(OPTIONAL) : Skip all test stages, but still run build stages, package stages and sanity check stages. Note: Does NOT update GitHub check status.--stage-list "A10-PyTorch-1, xxx"(OPTIONAL) : Only run the specified test stages. Examples: "A10-PyTorch-1, xxx". Note: Does NOT update GitHub check status.--gpu-type "A30, H100_PCIe"(OPTIONAL) : Only run the test stages on the specified GPU types. Examples: "A30, H100_PCIe". Note: Does NOT update GitHub check status.--test-backend "pytorch, cpp"(OPTIONAL) : Skip test stages which don't match the specified backends. Only support [pytorch, cpp, tensorrt, triton]. Examples: "pytorch, cpp" (does not run test stages with tensorrt or triton backend). Note: Does NOT update GitHub pipeline status.--only-multi-gpu-test(OPTIONAL) : Only run the multi-GPU tests. Note: Does NOT update GitHub check status.--disable-multi-gpu-test(OPTIONAL) : Disable the multi-GPU tests. Note: Does NOT update GitHub check status.--add-multi-gpu-test(OPTIONAL) : Force run the multi-GPU tests in addition to running L0 pre-merge pipeline.--post-merge(OPTIONAL) : Run the L0 post-merge pipeline instead of the ordinary L0 pre-merge pipeline.--extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx"(OPTIONAL) : Run the ordinary L0 pre-merge pipeline and specified test stages. Examples: --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx".--detailed-log(OPTIONAL) : Enable flushing out all logs to the Jenkins console. This will significantly increase the log volume and may slow down the job.--debug(OPTIONAL) : Experimental feature. Enable access to the CI container for debugging purpose. Note: Specify exactly one stage in thestage-listparameter to access the appropriate container environment. Note: Does NOT update GitHub check status.For guidance on mapping tests to stage names, see
docs/source/reference/ci-overview.mdand the
scripts/test_to_stage_mapping.pyhelper.kill
killKill all running builds associated with pull request.
skip
skip --comment COMMENTSkip testing for latest commit on pull request.
--comment "Reason for skipping build/test"is required. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.reuse-pipeline
reuse-pipelineReuse a previous pipeline to validate current commit. This action will also kill all currently running builds associated with the pull request. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.