[Observers] Small observers cleanup, add e2e quantization tests #1830

kylesayrs · 2025-09-17T13:29:02Z

Purpose

Add tests to quantization in order to add more confidence to future quantization refactors
- Each of the tested values has been checked and look correct

Changes

Observers changes
- Update calculate_updated_min_max to store running values, even if the function computation shortcuts (helpful testing)
- Update get_qparams_along_dim to support multiple dims and negative dims
  - This actually results in a silent typing bug with token quantization, and is fixed on the base class implementation
  - This change essentially duplicates the base class implementation. Future work could involve cleaning up the inheritance structure here

Testing

Added tests pass

Signed-off-by: Kyle Sayers <[email protected]>

github-actions · 2025-09-17T13:29:12Z

👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review.

Note: This is required to complete the testing suite, please only add the label once the PR is code complete and local testing has been performed.

gemini-code-assist

Summary of Changes

Hello @kylesayrs, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request primarily focuses on strengthening the quantization observer mechanisms and expanding test coverage. It refines the logic for calculating min-max values and global parameters within observers, ensuring more robust and predictable behavior. Crucially, it adds extensive end-to-end tests for different quantization strategies, which will serve as a vital safeguard for future modifications and improvements to the quantization framework.

Highlights

Observer Logic Refinements: Enhanced calculate_updated_min_max to consistently store running values and updated calculate_gparam to use patch_attr for safe, temporary state modification, preventing unintended side effects on running means.
Dimension Handling in Observers: Improved get_qparams_along_dim to support multiple and negative dimensions, increasing flexibility for quantization parameter calculation.
Comprehensive Quantization E2E Tests: Introduced new end-to-end tests for various quantization strategies (tensor, channel, group, block, token) covering both weight and activation quantization, significantly boosting confidence in the quantization pipeline.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	`/gemini review`	Performs a code review for the current pull request in its current state.
Pull Request Summary	`/gemini summary`	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	`/gemini help`	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

gemini-code-assist

Code Review

This pull request introduces valuable cleanups to the quantization observers and adds comprehensive end-to-end tests. The observer changes, such as making calculate_gparam side-effect-free and enhancing get_qparams_along_dim, are solid improvements. The new tests significantly increase confidence in the quantization logic for future refactoring. I've included a few suggestions to enhance the new test files by removing debugging print statements and commented-out code, which will improve overall maintainability. Great work on strengthening both the implementation and test coverage.

tests/llmcompressor/modifiers/calibration/test_observers.py

Signed-off-by: Kyle Sayers <[email protected]>

shanjiaz

Do you think we should update the mse observer as well? Since reset was removed from calibration.

kylesayrs · 2025-09-17T14:00:33Z

@shanjiaz global parameters are not supported by MSE observer anyways, so it's not relevant without that support

shanjiaz

Looks good to me!

Signed-off-by: Kyle Sayers <[email protected]>

src/llmcompressor/observers/min_max.py

brian-dellabetta · 2025-09-17T14:42:25Z

src/llmcompressor/observers/min_max.py

+        dim = set(dim)
+
+        # convert negative dims
+        dim = [d if d >= 0 else observed.ndim + d for d in dim]


Shouldn't the cast to set happen after this line?

Technically either is fine, since the argument type just needs to be an iterable. I'm purely matching the implementation on the base model for now

Update get_qparams_along_dim to support multiple dims and negative dims
This actually results in a silent typing bug with token quantization, and is fixed on the base class implementation
This change essentially duplicates the base class implementation. Future work could involve cleaning up the inheritance structure here

I mean more that you might end up with duplicates in dim if you create this list and don't cast back to a set.

e.g. if there are 3 dims and dim={1,2,-1}, then dim=[1,2,2] after this line.

brian-dellabetta · 2025-09-17T14:43:48Z

src/llmcompressor/observers/min_max.py

-        updated_min_val, updated_max_val = self.calculate_updated_min_max(
-            observed=observed
-        )
+        # patch to avoid affecting running means


this is because we are calculating global scale right? we don't want the calculate_qparams result to change based on this calculation?

Update calculate_gparam to restore original running values, rather than relying on resetting after calculation

Yes

Why is this preferable? If anything, this now seems more confusing

From a programming standpoint, this decouples calculate_gparam and Observer.reset (there's no way to footgun yourself by calling calculate_gparam and forgetting to call Observer.reset.

From a functionality standpoint, I think this fixes a bug where metrics would be updated twice (which has implications for running values), specifically when called from calibrate_activations. In the case of activations, we don't want to reset after each gparam calculation, since we still need those metrics to compute running values.

I think you're right about the 2nd point.

I don't know if I agree with the first point. This feels like a hack.

Case 1

Consider the case of strategy="tensor_group", dynamic=False and averaging_constant != 1.

On activation hook, calibrate_activations calls call_observer with should_calculate_gparam=True*. This causes calculate_updated_min_max to be called twice, which causes the running min/max to move faster than if no global param was calculated.

Case 2

Consider the case of strategy="tensor_group", dynamic="local" and averaging_constant != 1.

Originally, calculate_gparam would call calculate_updated_min_max would be called and the running values would update (twice*). Now, the running values will not update.

* Note that running values are updated, even if should_calculate_qparams=False

TLDR

So it seems that this change fixes a bug where running values are updated twice, but changes the behavior of dynamic="local" to calculate global parameters based on true values, not running values. I assumed that global parameters should be the true min/max of all values, not running values, but maybe @dsikka you think this shouldn't be the case?

I've reverted the change since it's not necessary for group quant, but we should definitely look into exactly the behavior we want for global scales (and scales in general. Runnings means are slightly strange anyways and seem to be a vestige of QAT).

dsikka · 2025-09-18T09:39:52Z

src/llmcompressor/modifiers/quantization/calibration.py

        should_calculate_gparam=True,
        should_calculate_qparams=False,
    )
-    module.weight_observer.reset()


Because we only attach one observer, I’m fairly sure we’re resetting to prevent global scale metrics from impacting quant scale metrics

Update calculate_gparam to restore original running values, rather than relying on resetting after calculation

Reset is replaced by patching and restoring the metrics

Co-authored-by: Brian Dellabetta <[email protected]>

Signed-off-by: Kyle Sayers <[email protected]>

kylesayrs · 2025-10-09T16:11:54Z

#1903

…servers (#1903) ## Purpose ## * FP4 * Fix bug discovered [here](#1830 (comment)) where dynamic="local" nvfp4 calculations would increment the observer twice as fast as normal * Enable MSE observer to be used with FP4 ```psuedocode mse_quant_error := mean((x - fake_quant(x))**2) global_scale <- min[min_vals, max_vals, global_scale](mse_quant_error(x)) scale, zp <- min[min_vals, max_vals](mse_quant_error(x, global_scale)) ``` * Simplification * Make supporting attention calibration easier by separating out weight/activation/attention reshaping * Improve readability of observer codes by removing many levels of function indirection * Drop support for calibration with non-divisible group sizes. This is not really a loss, since [forward passes](https://github.com/neuralmagic/compressed-tensors/blob/main/src/compressed_tensors/quantization/lifecycle/forward.py#L279) also make this assumption * New observers * `memoryless_minmax` computes min and max values on the fly in a dynamic-quantization style. This observer is useful for PTQ weight quantization * `static_minmax` computes absolute min and max values across all observations. This observer is useful for PTQ activation quantization * `memoryless_mse` computes best qparams w.r.t. MSE loss for each observation. This observer is useful for PTQ weight quantization * Memory improvements * All observers no longer store copies of scales and zero points, reducing the amount of required memory * Newly introduced "memoryless" observers do not store any quantization parameters, which greatly reduces the memory requirements for PTQ weight quantization of very large models | Diagrams | | - | | Before | | <img width="886" height="595" alt="before" src="https://github.com/user-attachments/assets/660d94c2-3ac8-4e05-9e9b-53d21145abac" /> | | After | <img width="1527" height="595" alt="after" src="https://github.com/user-attachments/assets/51a0107e-3fbd-413c-a7a6-03ddc3612169" /> | ## Changes ## * Standardize reshaping using `flatten_for_calibration` * This function reshapes all observed values to `(num_observations, *qparams_shape, group_size)` * This function the complexity associated with passing "reduce dims" and trying to handle weights, activations, and attention states all in the same function * In the future, this function could be applied to the quantization forward pass, although there's probably no need to outside of standardization * Implement `get_global_scale` on `Observer` base * This function decouples minmax calculations from regular qparam calculations (avoiding the double increment bug) * This function enables the MSE observer to be used with FP4 global scales ## Testing ## * Added additional minmax tests which check exact values of scales. This test passes both on main and this branch, demonstrating that minmax observer behavior remains unchanged * Added additional MSE tests which check exact values of mse losses. This test passes both on main and this branch, demonstrating that MSE observer behavior remains unchanged * Added FP4 MSE test ## Evaluation ## ``` nvfp4-static-minmax | Tasks |Version|Filter|n-shot| Metric | |Value | |Stderr| |--------|------:|------|-----:|--------|---|-----:|---|------| |mmmu_val| 0|none | 0|mmmu_acc|↑ |0.6167|± | N/A| ``` ``` nvfp4-minmax | Tasks |Version|Filter|n-shot| Metric | |Value | |Stderr| |--------|------:|------|-----:|--------|---|-----:|---|------| |mmmu_val| 0|none | 0|mmmu_acc|↑ |0.6011|± | N/A| ``` --------- Signed-off-by: Kyle Sayers <[email protected]> Signed-off-by: Dan Huang <[email protected]> Co-authored-by: dhuangnm <[email protected]>

…servers (#1903) * FP4 * Fix bug discovered [here](#1830 (comment)) where dynamic="local" nvfp4 calculations would increment the observer twice as fast as normal * Enable MSE observer to be used with FP4 ```psuedocode mse_quant_error := mean((x - fake_quant(x))**2) global_scale <- min[min_vals, max_vals, global_scale](mse_quant_error(x)) scale, zp <- min[min_vals, max_vals](mse_quant_error(x, global_scale)) ``` * Simplification * Make supporting attention calibration easier by separating out weight/activation/attention reshaping * Improve readability of observer codes by removing many levels of function indirection * Drop support for calibration with non-divisible group sizes. This is not really a loss, since [forward passes](https://github.com/neuralmagic/compressed-tensors/blob/main/src/compressed_tensors/quantization/lifecycle/forward.py#L279) also make this assumption * New observers * `memoryless_minmax` computes min and max values on the fly in a dynamic-quantization style. This observer is useful for PTQ weight quantization * `static_minmax` computes absolute min and max values across all observations. This observer is useful for PTQ activation quantization * `memoryless_mse` computes best qparams w.r.t. MSE loss for each observation. This observer is useful for PTQ weight quantization * Memory improvements * All observers no longer store copies of scales and zero points, reducing the amount of required memory * Newly introduced "memoryless" observers do not store any quantization parameters, which greatly reduces the memory requirements for PTQ weight quantization of very large models | Diagrams | | - | | Before | | <img width="886" height="595" alt="before" src="https://github.com/user-attachments/assets/660d94c2-3ac8-4e05-9e9b-53d21145abac" /> | | After | <img width="1527" height="595" alt="after" src="https://github.com/user-attachments/assets/51a0107e-3fbd-413c-a7a6-03ddc3612169" /> | * Standardize reshaping using `flatten_for_calibration` * This function reshapes all observed values to `(num_observations, *qparams_shape, group_size)` * This function the complexity associated with passing "reduce dims" and trying to handle weights, activations, and attention states all in the same function * In the future, this function could be applied to the quantization forward pass, although there's probably no need to outside of standardization * Implement `get_global_scale` on `Observer` base * This function decouples minmax calculations from regular qparam calculations (avoiding the double increment bug) * This function enables the MSE observer to be used with FP4 global scales * Added additional minmax tests which check exact values of scales. This test passes both on main and this branch, demonstrating that minmax observer behavior remains unchanged * Added additional MSE tests which check exact values of mse losses. This test passes both on main and this branch, demonstrating that MSE observer behavior remains unchanged * Added FP4 MSE test ``` nvfp4-static-minmax | Tasks |Version|Filter|n-shot| Metric | |Value | |Stderr| |--------|------:|------|-----:|--------|---|-----:|---|------| |mmmu_val| 0|none | 0|mmmu_acc|↑ |0.6167|± | N/A| ``` ``` nvfp4-minmax | Tasks |Version|Filter|n-shot| Metric | |Value | |Stderr| |--------|------:|------|-----:|--------|---|-----:|---|------| |mmmu_val| 0|none | 0|mmmu_acc|↑ |0.6011|± | N/A| ``` --------- Signed-off-by: Kyle Sayers <[email protected]> Signed-off-by: Dan Huang <[email protected]> Co-authored-by: dhuangnm <[email protected]>

…servers (vllm-project#1903) ## Purpose ## * FP4 * Fix bug discovered [here](vllm-project#1830 (comment)) where dynamic="local" nvfp4 calculations would increment the observer twice as fast as normal * Enable MSE observer to be used with FP4 ```psuedocode mse_quant_error := mean((x - fake_quant(x))**2) global_scale <- min[min_vals, max_vals, global_scale](mse_quant_error(x)) scale, zp <- min[min_vals, max_vals](mse_quant_error(x, global_scale)) ``` * Simplification * Make supporting attention calibration easier by separating out weight/activation/attention reshaping * Improve readability of observer codes by removing many levels of function indirection * Drop support for calibration with non-divisible group sizes. This is not really a loss, since [forward passes](https://github.com/neuralmagic/compressed-tensors/blob/main/src/compressed_tensors/quantization/lifecycle/forward.py#L279) also make this assumption * New observers * `memoryless_minmax` computes min and max values on the fly in a dynamic-quantization style. This observer is useful for PTQ weight quantization * `static_minmax` computes absolute min and max values across all observations. This observer is useful for PTQ activation quantization * `memoryless_mse` computes best qparams w.r.t. MSE loss for each observation. This observer is useful for PTQ weight quantization * Memory improvements * All observers no longer store copies of scales and zero points, reducing the amount of required memory * Newly introduced "memoryless" observers do not store any quantization parameters, which greatly reduces the memory requirements for PTQ weight quantization of very large models | Diagrams | | - | | Before | | <img width="886" height="595" alt="before" src="https://github.com/user-attachments/assets/660d94c2-3ac8-4e05-9e9b-53d21145abac" /> | | After | <img width="1527" height="595" alt="after" src="https://github.com/user-attachments/assets/51a0107e-3fbd-413c-a7a6-03ddc3612169" /> | ## Changes ## * Standardize reshaping using `flatten_for_calibration` * This function reshapes all observed values to `(num_observations, *qparams_shape, group_size)` * This function the complexity associated with passing "reduce dims" and trying to handle weights, activations, and attention states all in the same function * In the future, this function could be applied to the quantization forward pass, although there's probably no need to outside of standardization * Implement `get_global_scale` on `Observer` base * This function decouples minmax calculations from regular qparam calculations (avoiding the double increment bug) * This function enables the MSE observer to be used with FP4 global scales ## Testing ## * Added additional minmax tests which check exact values of scales. This test passes both on main and this branch, demonstrating that minmax observer behavior remains unchanged * Added additional MSE tests which check exact values of mse losses. This test passes both on main and this branch, demonstrating that MSE observer behavior remains unchanged * Added FP4 MSE test ## Evaluation ## ``` nvfp4-static-minmax | Tasks |Version|Filter|n-shot| Metric | |Value | |Stderr| |--------|------:|------|-----:|--------|---|-----:|---|------| |mmmu_val| 0|none | 0|mmmu_acc|↑ |0.6167|± | N/A| ``` ``` nvfp4-minmax | Tasks |Version|Filter|n-shot| Metric | |Value | |Stderr| |--------|------:|------|-----:|--------|---|-----:|---|------| |mmmu_val| 0|none | 0|mmmu_acc|↑ |0.6011|± | N/A| ``` --------- Signed-off-by: Kyle Sayers <[email protected]> Signed-off-by: Dan Huang <[email protected]> Co-authored-by: dhuangnm <[email protected]> Signed-off-by: ronantakizawa <[email protected]>

…servers (vllm-project#1903) ## Purpose ## * FP4 * Fix bug discovered [here](vllm-project#1830 (comment)) where dynamic="local" nvfp4 calculations would increment the observer twice as fast as normal * Enable MSE observer to be used with FP4 ```psuedocode mse_quant_error := mean((x - fake_quant(x))**2) global_scale <- min[min_vals, max_vals, global_scale](mse_quant_error(x)) scale, zp <- min[min_vals, max_vals](mse_quant_error(x, global_scale)) ``` * Simplification * Make supporting attention calibration easier by separating out weight/activation/attention reshaping * Improve readability of observer codes by removing many levels of function indirection * Drop support for calibration with non-divisible group sizes. This is not really a loss, since [forward passes](https://github.com/neuralmagic/compressed-tensors/blob/main/src/compressed_tensors/quantization/lifecycle/forward.py#L279) also make this assumption * New observers * `memoryless_minmax` computes min and max values on the fly in a dynamic-quantization style. This observer is useful for PTQ weight quantization * `static_minmax` computes absolute min and max values across all observations. This observer is useful for PTQ activation quantization * `memoryless_mse` computes best qparams w.r.t. MSE loss for each observation. This observer is useful for PTQ weight quantization * Memory improvements * All observers no longer store copies of scales and zero points, reducing the amount of required memory * Newly introduced "memoryless" observers do not store any quantization parameters, which greatly reduces the memory requirements for PTQ weight quantization of very large models | Diagrams | | - | | Before | | <img width="886" height="595" alt="before" src="https://github.com/user-attachments/assets/660d94c2-3ac8-4e05-9e9b-53d21145abac" /> | | After | <img width="1527" height="595" alt="after" src="https://github.com/user-attachments/assets/51a0107e-3fbd-413c-a7a6-03ddc3612169" /> | ## Changes ## * Standardize reshaping using `flatten_for_calibration` * This function reshapes all observed values to `(num_observations, *qparams_shape, group_size)` * This function the complexity associated with passing "reduce dims" and trying to handle weights, activations, and attention states all in the same function * In the future, this function could be applied to the quantization forward pass, although there's probably no need to outside of standardization * Implement `get_global_scale` on `Observer` base * This function decouples minmax calculations from regular qparam calculations (avoiding the double increment bug) * This function enables the MSE observer to be used with FP4 global scales ## Testing ## * Added additional minmax tests which check exact values of scales. This test passes both on main and this branch, demonstrating that minmax observer behavior remains unchanged * Added additional MSE tests which check exact values of mse losses. This test passes both on main and this branch, demonstrating that MSE observer behavior remains unchanged * Added FP4 MSE test ## Evaluation ## ``` nvfp4-static-minmax | Tasks |Version|Filter|n-shot| Metric | |Value | |Stderr| |--------|------:|------|-----:|--------|---|-----:|---|------| |mmmu_val| 0|none | 0|mmmu_acc|↑ |0.6167|± | N/A| ``` ``` nvfp4-minmax | Tasks |Version|Filter|n-shot| Metric | |Value | |Stderr| |--------|------:|------|-----:|--------|---|-----:|---|------| |mmmu_val| 0|none | 0|mmmu_acc|↑ |0.6011|± | N/A| ``` --------- Signed-off-by: Kyle Sayers <[email protected]> Signed-off-by: Dan Huang <[email protected]> Co-authored-by: dhuangnm <[email protected]>

…servers (vllm-project#1903) ## Purpose ## * FP4 * Fix bug discovered [here](vllm-project#1830 (comment)) where dynamic="local" nvfp4 calculations would increment the observer twice as fast as normal * Enable MSE observer to be used with FP4 ```psuedocode mse_quant_error := mean((x - fake_quant(x))**2) global_scale <- min[min_vals, max_vals, global_scale](mse_quant_error(x)) scale, zp <- min[min_vals, max_vals](mse_quant_error(x, global_scale)) ``` * Simplification * Make supporting attention calibration easier by separating out weight/activation/attention reshaping * Improve readability of observer codes by removing many levels of function indirection * Drop support for calibration with non-divisible group sizes. This is not really a loss, since [forward passes](https://github.com/neuralmagic/compressed-tensors/blob/main/src/compressed_tensors/quantization/lifecycle/forward.py#L279) also make this assumption * New observers * `memoryless_minmax` computes min and max values on the fly in a dynamic-quantization style. This observer is useful for PTQ weight quantization * `static_minmax` computes absolute min and max values across all observations. This observer is useful for PTQ activation quantization * `memoryless_mse` computes best qparams w.r.t. MSE loss for each observation. This observer is useful for PTQ weight quantization * Memory improvements * All observers no longer store copies of scales and zero points, reducing the amount of required memory * Newly introduced "memoryless" observers do not store any quantization parameters, which greatly reduces the memory requirements for PTQ weight quantization of very large models | Diagrams | | - | | Before | | <img width="886" height="595" alt="before" src="https://github.com/user-attachments/assets/660d94c2-3ac8-4e05-9e9b-53d21145abac" /> | | After | <img width="1527" height="595" alt="after" src="https://github.com/user-attachments/assets/51a0107e-3fbd-413c-a7a6-03ddc3612169" /> | ## Changes ## * Standardize reshaping using `flatten_for_calibration` * This function reshapes all observed values to `(num_observations, *qparams_shape, group_size)` * This function the complexity associated with passing "reduce dims" and trying to handle weights, activations, and attention states all in the same function * In the future, this function could be applied to the quantization forward pass, although there's probably no need to outside of standardization * Implement `get_global_scale` on `Observer` base * This function decouples minmax calculations from regular qparam calculations (avoiding the double increment bug) * This function enables the MSE observer to be used with FP4 global scales ## Testing ## * Added additional minmax tests which check exact values of scales. This test passes both on main and this branch, demonstrating that minmax observer behavior remains unchanged * Added additional MSE tests which check exact values of mse losses. This test passes both on main and this branch, demonstrating that MSE observer behavior remains unchanged * Added FP4 MSE test ## Evaluation ## ``` nvfp4-static-minmax | Tasks |Version|Filter|n-shot| Metric | |Value | |Stderr| |--------|------:|------|-----:|--------|---|-----:|---|------| |mmmu_val| 0|none | 0|mmmu_acc|↑ |0.6167|± | N/A| ``` ``` nvfp4-minmax | Tasks |Version|Filter|n-shot| Metric | |Value | |Stderr| |--------|------:|------|-----:|--------|---|-----:|---|------| |mmmu_val| 0|none | 0|mmmu_acc|↑ |0.6011|± | N/A| ``` --------- Signed-off-by: Kyle Sayers <[email protected]> Signed-off-by: Dan Huang <[email protected]> Co-authored-by: dhuangnm <[email protected]> Signed-off-by: LeiZhang <[email protected]>

kylesayrs added 2 commits September 17, 2025 09:20

add tests

7a3d43a

Signed-off-by: Kyle Sayers <[email protected]>

remove reset

25e54a5

Signed-off-by: Kyle Sayers <[email protected]>

gemini-code-assist bot reviewed Sep 17, 2025

View reviewed changes

kylesayrs mentioned this pull request Sep 17, 2025

[Quantization] Group Activation Quantization #1811

Closed

gemini-code-assist bot reviewed Sep 17, 2025

View reviewed changes

tests/llmcompressor/modifiers/calibration/test_observers.py Outdated Show resolved Hide resolved

tests/llmcompressor/modifiers/calibration/test_observers.py Show resolved Hide resolved

tests/llmcompressor/modifiers/calibration/test_observers.py Outdated Show resolved Hide resolved

kylesayrs added 2 commits September 17, 2025 09:31

remove prints

06ece86

Signed-off-by: Kyle Sayers <[email protected]>

remove commented code

178d0ae

Signed-off-by: Kyle Sayers <[email protected]>

shanjiaz reviewed Sep 17, 2025

View reviewed changes

shanjiaz previously approved these changes Sep 17, 2025

View reviewed changes

increase limit

7d81a79

Signed-off-by: Kyle Sayers <[email protected]>

kylesayrs dismissed shanjiaz’s stale review via 7d81a79 September 17, 2025 14:03

kylesayrs changed the title ~~[Observers] Small observers cleanup, add e2e quantization tests~~ [Observers] Small observers cleanup, add e2e quantization tests, increase test file limit Sep 17, 2025

brian-dellabetta reviewed Sep 17, 2025

View reviewed changes

dsikka reviewed Sep 18, 2025

View reviewed changes

kylesayrs and others added 2 commits September 18, 2025 10:45

Update src/llmcompressor/observers/min_max.py

05d780a

Co-authored-by: Brian Dellabetta <[email protected]>

revert global disjointness

dd91329

Signed-off-by: Kyle Sayers <[email protected]>

kylesayrs changed the title ~~[Observers] Small observers cleanup, add e2e quantization tests, increase test file limit~~ [Observers] Small observers cleanup, add e2e quantization tests Sep 18, 2025

kylesayrs added 2 commits September 18, 2025 07:54

remove safe permute

a6b3842

Signed-off-by: Kyle Sayers <[email protected]>

fix typo

c098447

Signed-off-by: Kyle Sayers <[email protected]>

kylesayrs mentioned this pull request Oct 7, 2025

[Observers] Refactor for better FP4 support, static and memoryless observers #1903

Merged

kylesayrs closed this Oct 9, 2025

Uh oh!

[Observers] Small observers cleanup, add e2e quantization tests #1830

[Observers] Small observers cleanup, add e2e quantization tests #1830

Conversation

kylesayrs commented Sep 17, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Purpose

Changes

Testing

Uh oh!

github-actions bot commented Sep 17, 2025

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Summary of Changes

Highlights

Footnotes

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

Uh oh!

Uh oh!

Uh oh!

shanjiaz left a comment

Choose a reason for hiding this comment

Uh oh!

kylesayrs commented Sep 17, 2025

Uh oh!

shanjiaz left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Case 1

Case 2

TLDR

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

kylesayrs commented Oct 9, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

kylesayrs commented Sep 17, 2025 •

edited

Loading