enable finegrained_fp8 and granite_speech cases on XPU #38036

yao-matrix · 2025-05-09T07:30:12Z

@ydshieh @IlyasMoutawwakil , pls help review, thx

Signed-off-by: Yao Matrix <[email protected]>

github-actions · 2025-05-09T07:30:25Z

Hi 👋, thank you for opening this pull request! The pull request is converted to draft by default. The CI will be paused while the PR is in draft mode. When it is ready for review, please click the Ready for review button (at the bottom of the PR page). This will assign reviewers and trigger CI.

Signed-off-by: Yao Matrix <[email protected]>

src/transformers/quantizers/quantizer_finegrained_fp8.py

yao-matrix · 2025-05-09T07:43:23Z

tests/quantization/finegrained_fp8/test_fp8.py

@@ -188,11 +191,12 @@ def test_block_size(self):
        )
        self.assertEqual(quantized_model.config.quantization_config.weight_block_size, (32, 32))

-    @require_torch_multi_gpu
-    def test_quantized_model_multi_gpu(self):


For multi_accelerator case ground truth. I can find both XPU and A100 return device 0, but the ground truth is {0,1}.
I look a bit deeper, for the target model meta-llama/Llama-3.2-1B, actually it has 2 modules_to_treat in accelerate infer_auto_device_map, as below. and the lm_head is tied with model, which means it only has 1 module to treat, and naturally can only placed to device 0.

I don't know whether there are any other scenarios I didn't considered, but for this case, seems the correct ground truth should be 0. @ydshieh , pls let me know your insights, thx

[('model', LlamaModel(
(embed_tokens): Embedding(128256, 2048)
(layers): ModuleList(
(0-15): 16 x LlamaDecoderLayer(
(self_attn): LlamaAttention(
(q_proj): FP8Linear(in_features=2048, out_features=2048, bias=False)
(k_proj): FP8Linear(in_features=2048, out_features=512, bias=False)
(v_proj): FP8Linear(in_features=2048, out_features=512, bias=False)
(o_proj): FP8Linear(in_features=2048, out_features=2048, bias=False)
)
(mlp): LlamaMLP(
(gate_proj): FP8Linear(in_features=2048, out_features=8192, bias=False)
(up_proj): FP8Linear(in_features=2048, out_features=8192, bias=False)
(down_proj): FP8Linear(in_features=8192, out_features=2048, bias=False)
(act_fn): SiLU()
)
(input_layernorm): LlamaRMSNorm((2048,), eps=1e-05)
(post_attention_layernorm): LlamaRMSNorm((2048,), eps=1e-05)
)
)
(norm): LlamaRMSNorm((2048,), eps=1e-05)
(rotary_emb): LlamaRotaryEmbedding()
)), ('lm_head', Linear(in_features=2048, out_features=128256, bias=False))]

I believe this depends on the vram of your GPUs/XPUs ; it will only use both if one is not enough, otherwise maybe it would make sense to use another device map strategy here lik "balanced"

I will let @SunMarc or @MekkCyber to share their thoughts for this.

On our CI, these tests are not collected, I believe it is due to the require_read_token decorator at the class level.

@yao-matrix You are able to run this test ...? I am surprised. I will take a look at this issue

embed_tokens is indeed tied to lm_head but the layers can be dispatched to other gpus. setting "auto" in device_map will default to "balanced" strategy.

I will let @SunMarc or @MekkCyber to share their thoughts for this.

On our CI, these tests are not collected, I believe it is due to the require_read_token decorator at the class level.

@yao-matrix You are able to run this test ...? I am surprised. I will take a look at this issue

I removed require_read_token in my local env and run this case.

@IlyasMoutawwakil @SunMarc yes, i tried balanced in my local env too w/ the same consideration as yours(my XPU has 64GB VRAM), but the result is still 1. It seems split granularity is top-level module when available memory is enough to fit it, in infer_auto_device_map

Will investigate a bit more. @MekkCyber tested locally and it works but when running with pytest it fails

@SunMarc @MekkCyber You will need to remove require_read_token decorator in order to run these tests.
(if not done so yet)

related issue: #38093

If you think there is no more change required for this test_quantized_model_multi_gpu, feel free give a ✅ 🙏 .

From my side, I am just waiting a nit change regarding variable name.

I think it's good to go from my side, we need to figure out why it fails with pytest but no need to include that in this pr, and thanks @ydshieh for the advice about require_read_token

src/transformers/integrations/finegrained_fp8.py

tests/quantization/finegrained_fp8/test_fp8.py

SunMarc

Thanks ! Left some comments

src/transformers/quantizers/quantizer_finegrained_fp8.py

SunMarc · 2025-05-12T14:54:53Z

tests/quantization/finegrained_fp8/test_fp8.py

@@ -188,11 +191,12 @@ def test_block_size(self):
        )
        self.assertEqual(quantized_model.config.quantization_config.weight_block_size, (32, 32))

-    @require_torch_multi_gpu
-    def test_quantized_model_multi_gpu(self):


embed_tokens is indeed tied to lm_head but the layers can be dispatched to other gpus. setting "auto" in device_map will default to "balanced" strategy.

tests/quantization/finegrained_fp8/test_fp8.py

Signed-off-by: Matrix Yao <[email protected]>

yao-matrix · 2025-05-13T22:48:29Z

@ydshieh , resolved comments, thx

MekkCyber

Thanks 🤗

HuggingFaceDocBuilderDev · 2025-05-14T08:59:12Z

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

…8036) * enable finegrained_fp8 cases on XPU Signed-off-by: Yao Matrix <[email protected]> * fix style Signed-off-by: Yao Matrix <[email protected]> * change back to auto Signed-off-by: Yao Matrix <[email protected]> * rename per comments Signed-off-by: Matrix Yao <[email protected]> --------- Signed-off-by: Yao Matrix <[email protected]> Signed-off-by: Matrix Yao <[email protected]> Co-authored-by: Marc Sun <[email protected]>

yao-matrix added 3 commits May 9, 2025 07:26

enable finegrained_fp8 cases on XPU

c2b7d1a

Signed-off-by: Yao Matrix <[email protected]>

fix style

ab89c7c

Signed-off-by: Yao Matrix <[email protected]>

Merge branch 'main' into xpu

234b759

github-actions bot marked this pull request as draft May 9, 2025 07:30

change back to auto

fbb1c67

Signed-off-by: Yao Matrix <[email protected]>

yao-matrix commented May 9, 2025

View reviewed changes

yao-matrix marked this pull request as ready for review May 9, 2025 07:43

IlyasMoutawwakil requested a review from MekkCyber May 9, 2025 10:29

Merge branch 'main' into xpu

612e51d

ydshieh reviewed May 12, 2025

View reviewed changes

src/transformers/integrations/finegrained_fp8.py Show resolved Hide resolved

ydshieh reviewed May 12, 2025

View reviewed changes

tests/quantization/finegrained_fp8/test_fp8.py Outdated Show resolved Hide resolved

SunMarc reviewed May 12, 2025

View reviewed changes

yao-matrix and others added 4 commits May 13, 2025 14:00

Merge branch 'main' into xpu

7dded96

Merge branch 'main' into xpu

56639cf

Merge branch 'main' into xpu

fe4ce7b

rename per comments

279faa9

Signed-off-by: Matrix Yao <[email protected]>

MekkCyber approved these changes May 14, 2025

View reviewed changes

ydshieh enabled auto-merge (squash) May 14, 2025 08:46

ydshieh merged commit 9b5ce55 into huggingface:main May 14, 2025
20 checks passed

yao-matrix deleted the xpu branch May 14, 2025 23:02

enable finegrained_fp8 and granite_speech cases on XPU #38036

enable finegrained_fp8 and granite_speech cases on XPU #38036

Uh oh!

Conversation

yao-matrix commented May 9, 2025

Uh oh!

github-actions bot commented May 9, 2025

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

IlyasMoutawwakil May 12, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

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!

yao-matrix May 12, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SunMarc May 13, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

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!

Uh oh!

Uh oh!

SunMarc left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

yao-matrix commented May 13, 2025

Uh oh!

MekkCyber left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

HuggingFaceDocBuilderDev commented May 14, 2025

Uh oh!

Uh oh!

IlyasMoutawwakil May 12, 2025 •

edited

Loading

yao-matrix May 12, 2025 •

edited

Loading

SunMarc May 13, 2025 •

edited

Loading