Skip to content

Commit

Permalink
Tests: remove cuda versions when the result is the same 🧹🧹 (huggingfa…
Browse files Browse the repository at this point in the history 
…ce#31955)

remove cuda versions when the result is the same
  • Loading branch information
@gante @zucchini-nlp
gante authored and zucchini-nlp committed Jul 24, 2024
1 parent 0cfaedf commit 47e9238
Show file tree
Hide file tree
Showing 3 changed files with 52 additions and 155 deletions.
123 changes: 29 additions & 94 deletions tests/models/gemma/test_modeling_gemma.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -566,24 +566,10 @@ def test_model_2b_fp16(self):
def test_model_2b_bf16(self):
model_id = "google/gemma-2b"

# Key 9 for MI300, Key 8 for A100/A10, and Key 7 for T4.
#
# Note: Key 9 is currently set for MI300, but may need potential future adjustments for H100s,
# considering differences in hardware processing and potential deviations in generated text.
EXPECTED_TEXTS = {
7: [
"Hello I am doing a project on the 1990s and I need to know what the most popular music",
"Hi today I am going to share with you a very easy and simple recipe of <strong><em>Khichdi",
],
8: [
"Hello I am doing a project on the 1990s and I need to know what the most popular music",
"Hi today I am going to share with you a very easy and simple recipe of <strong><em>Kaju Kat",
],
9: [
"Hello I am doing a project on the 1990s and I need to know what the most popular music",
"Hi today I am going to share with you a very easy and simple recipe of <strong><em>Kaju Kat",
],
}
EXPECTED_TEXTS = [
"Hello I am doing a project on the 1990s and I need to know what the most popular music",
"Hi today I am going to share with you a very easy and simple recipe of <strong><em>Khichdi",
]

model = AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True, torch_dtype=torch.bfloat16).to(
torch_device
Expand All @@ -595,30 +581,16 @@ def test_model_2b_bf16(self):
output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
output_text = tokenizer.batch_decode(output, skip_special_tokens=True)

self.assertEqual(output_text, EXPECTED_TEXTS[self.cuda_compute_capability_major_version])
self.assertEqual(output_text, EXPECTED_TEXTS)

@require_read_token
def test_model_2b_eager(self):
model_id = "google/gemma-2b"

# Key 9 for MI300, Key 8 for A100/A10, and Key 7 for T4.
#
# Note: Key 9 is currently set for MI300, but may need potential future adjustments for H100s,
# considering differences in hardware processing and potential deviations in generated text.
EXPECTED_TEXTS = {
7: [
"Hello I am doing a project on the 1990s and I need to know what the most popular music",
"Hi today I am going to share with you a very easy and simple recipe of <strong><em>Khichdi",
],
8: [
"Hello I am doing a project on the 1990s and I need to know what the most popular music",
"Hi today I am going to share with you a very easy and simple recipe of <strong><em>Kaju Kat",
],
9: [
"Hello I am doing a project on the 1990s and I need to know what the most popular music",
"Hi today I am going to share with you a very easy and simple recipe of <strong><em>Kaju Kat",
],
}
EXPECTED_TEXTS = [
"Hello I am doing a project on the 1990s and I need to know what the most popular music",
"Hi today I am going to share with you a very easy and simple recipe of <strong><em>Khichdi",
]

model = AutoModelForCausalLM.from_pretrained(
model_id, low_cpu_mem_usage=True, torch_dtype=torch.bfloat16, attn_implementation="eager"
Expand All @@ -631,31 +603,17 @@ def test_model_2b_eager(self):
output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
output_text = tokenizer.batch_decode(output, skip_special_tokens=True)

self.assertEqual(output_text, EXPECTED_TEXTS[self.cuda_compute_capability_major_version])
self.assertEqual(output_text, EXPECTED_TEXTS)

@require_torch_sdpa
@require_read_token
def test_model_2b_sdpa(self):
model_id = "google/gemma-2b"

# Key 9 for MI300, Key 8 for A100/A10, and Key 7 for T4.
#
# Note: Key 9 is currently set for MI300, but may need potential future adjustments for H100s,
# considering differences in hardware processing and potential deviations in generated text.
EXPECTED_TEXTS = {
7: [
"Hello I am doing a project on the 1990s and I need to know what the most popular music",
"Hi today I am going to share with you a very easy and simple recipe of <strong><em>Khichdi",
],
8: [
"Hello I am doing a project on the 1990s and I need to know what the most popular music",
"Hi today I am going to share with you a very easy and simple recipe of <strong><em>Kaju Kat",
],
9: [
"Hello I am doing a project on the 1990s and I need to know what the most popular music",
"Hi today I am going to share with you a very easy and simple recipe of <strong><em>Kaju Kat",
],
}
EXPECTED_TEXTS = [
"Hello I am doing a project on the 1990s and I need to know what the most popular music",
"Hi today I am going to share with you a very easy and simple recipe of <strong><em>Khichdi",
]

model = AutoModelForCausalLM.from_pretrained(
model_id, low_cpu_mem_usage=True, torch_dtype=torch.bfloat16, attn_implementation="sdpa"
Expand All @@ -668,7 +626,7 @@ def test_model_2b_sdpa(self):
output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
output_text = tokenizer.batch_decode(output, skip_special_tokens=True)

self.assertEqual(output_text, EXPECTED_TEXTS[self.cuda_compute_capability_major_version])
self.assertEqual(output_text, EXPECTED_TEXTS)

@pytest.mark.flash_attn_test
@require_flash_attn
Expand Down Expand Up @@ -734,7 +692,7 @@ def test_model_7b_fp32(self):
@require_read_token
def test_model_7b_fp16(self):
if self.cuda_compute_capability_major_version == 7:
self.skipTest("This test is failing (`torch.compile` fails) on Nvidia T4 GPU.")
self.skipTest("This test is failing (`torch.compile` fails) on Nvidia T4 GPU (OOM).")

model_id = "google/gemma-7b"
EXPECTED_TEXTS = [
Expand All @@ -757,7 +715,7 @@ def test_model_7b_fp16(self):
@require_read_token
def test_model_7b_bf16(self):
if self.cuda_compute_capability_major_version == 7:
self.skipTest("This test is failing (`torch.compile` fails) on Nvidia T4 GPU.")
self.skipTest("This test is failing (`torch.compile` fails) on Nvidia T4 GPU (OOM).")

model_id = "google/gemma-7b"

Expand Down Expand Up @@ -795,7 +753,7 @@ def test_model_7b_bf16(self):
@require_read_token
def test_model_7b_fp16_static_cache(self):
if self.cuda_compute_capability_major_version == 7:
self.skipTest("This test is failing (`torch.compile` fails) on Nvidia T4 GPU.")
self.skipTest("This test is failing (`torch.compile` fails) on Nvidia T4 GPU (OOM).")

model_id = "google/gemma-7b"
EXPECTED_TEXTS = [
Expand All @@ -821,16 +779,10 @@ def test_model_7b_fp16_static_cache(self):
@require_read_token
def test_model_7b_4bit(self):
model_id = "google/gemma-7b"
EXPECTED_TEXTS = {
7: [
"Hello I am doing a project for my school and I am trying to make a program that will take a number and then",
"Hi today I am going to talk about the best way to get rid of acne. miniaturing is a very",
],
8: [
"Hello I am doing a project for my school and I am trying to make a program that will take a number and then",
"Hi today I am going to talk about the best way to get rid of acne. miniaturing is a very",
],
}
EXPECTED_TEXTS = [
"Hello I am doing a project for my school and I am trying to make a program that will take a number and then",
"Hi today I am going to talk about the best way to get rid of acne. miniaturing is a very",
]

model = AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True, load_in_4bit=True)

Expand All @@ -839,7 +791,7 @@ def test_model_7b_4bit(self):

output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
output_text = tokenizer.batch_decode(output, skip_special_tokens=True)
self.assertEqual(output_text, EXPECTED_TEXTS[self.cuda_compute_capability_major_version])
self.assertEqual(output_text, EXPECTED_TEXTS)

@slow
@require_torch_gpu
Expand All @@ -851,27 +803,10 @@ def test_compile_static_cache(self):
self.skipTest(reason="This test requires torch >= 2.3 to run.")

NUM_TOKENS_TO_GENERATE = 40
# Note on `EXPECTED_TEXT_COMPLETION`'s diff: the current value matches the original test if the original test
# was changed to have a cache of 53 tokens (as opposed to 4096), on Ampere GPUs.
#
# Key 9 for MI300, Key 8 for A100/A10, and Key 7 for T4.
#
# Note: Key 9 is currently set for MI300, but may need potential future adjustments for H100s,
# considering differences in hardware processing and potential deviations in generated text.
EXPECTED_TEXT_COMPLETION = {
8: [
"Hello I am doing a project on the 1990s and I need to know what the most popular music was in the 1990s. I have looked on the internet and I have found",
"Hi today\nI have a problem with my 2007 1.9 tdi 105bhp.\nI have a problem with the engine management light on.\nI have checked the",
],
7: [
"Hello I am doing a project on the 1990s and I need to know what the most popular music was in the 1990s. I have looked on the internet and I have found",
"Hi today\nI have a problem with my 2007 1.9 tdi 105bhp.\nI have a problem with the engine management light on.\nI have checked the",
],
9: [
"Hello I am doing a project on the 1990s and I need to know what the most popular music was in the 1990s. I have looked on the internet and I have found",
"Hi today\nI have a problem with my 2007 1.9 tdi 105bhp.\nI have a problem with the engine management light on.\nI have checked the",
],
}
EXPECTED_TEXT_COMPLETION = [
"Hello I am doing a project on the 1990s and I need to know what the most popular music was in the 1990s. I have looked on the internet and I have found",
"Hi today\nI have a problem with my 2007 1.9 tdi 105bhp.\nI have a problem with the engine management light on.\nI have checked the",
]

prompts = ["Hello I am doing", "Hi today"]
tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b", pad_token="</s>", padding_side="right")
Expand All @@ -888,15 +823,15 @@ def test_compile_static_cache(self):
**inputs, max_new_tokens=NUM_TOKENS_TO_GENERATE, do_sample=False, cache_implementation="static"
)
static_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
self.assertEqual(EXPECTED_TEXT_COMPLETION[self.cuda_compute_capability_major_version], static_text)
self.assertEqual(EXPECTED_TEXT_COMPLETION, static_text)

# Static Cache + compile
model.forward = torch.compile(model.forward, mode="reduce-overhead", fullgraph=True)
generated_ids = model.generate(
**inputs, max_new_tokens=NUM_TOKENS_TO_GENERATE, do_sample=False, cache_implementation="static"
)
static_compiled_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
self.assertEqual(EXPECTED_TEXT_COMPLETION[self.cuda_compute_capability_major_version], static_compiled_text)
self.assertEqual(EXPECTED_TEXT_COMPLETION, static_compiled_text)

def test_model_2b_bf16_dola(self):
model_id = "google/gemma-2b"
Expand Down
41 changes: 10 additions & 31 deletions tests/models/llama/test_modeling_llama.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -738,32 +738,13 @@ def test_compile_static_cache(self):
NUM_TOKENS_TO_GENERATE = 40
# Note on `EXPECTED_TEXT_COMPLETION`'s diff: the current value matches the original test if the original test
# was changed to have a cache of 53 tokens (as opposed to 4096), on Ampere GPUs.
#
# Key 9 for MI300, Key 8 for A100/A10, and Key 7 for T4.
#
# Note: Key 9 is currently set for MI300, but may need potential future adjustments for H100s,
# considering differences in hardware processing and potential deviations in generated text.
EXPECTED_TEXT_COMPLETION = {
8: [
"Simply put, the theory of relativity states that 1) the speed of light is constant in all inertial "
"reference frames, and 2) the laws of physics are the same for all inertial reference frames.\nThe "
"theory of relativ",
"My favorite all time favorite condiment is ketchup. I love it on everything. I love it on my eggs, "
"my fries, my chicken, my burgers, my hot dogs, my sandwiches, my salads, my p",
],
7: [
"Simply put, the theory of relativity states that 1) the speed of light is constant in all inertial reference frames, and 2) the laws of physics are the same for all inertial reference frames.\nThe theory of relativ",
"My favorite all time favorite condiment is ketchup. I love it on everything. I love it on my eggs, my fries, my chicken, my burgers, my hot dogs, my sandwiches, my salads, my p",
],
9: [
"Simply put, the theory of relativity states that 1) the speed of light is constant in all inertial"
" reference frames, and 2) the laws of physics are the same for all inertial reference frames.\nThe "
"theory of relativ",
"My favorite all time favorite condiment is ketchup. I love it on everything. I love it on my eggs,"
" my fries, my chicken, my burgers, my hot dogs, my sandwiches, my salads, my p",
],
}
expected_text_completion_idx = 8
EXPECTED_TEXT_COMPLETION = [
"Simply put, the theory of relativity states that 1) the speed of light is constant in all inertial "
"reference frames, and 2) the laws of physics are the same for all inertial reference frames.\nThe "
"theory of relativ",
"My favorite all time favorite condiment is ketchup. I love it on everything. I love it on my eggs, "
"my fries, my chicken, my burgers, my hot dogs, my sandwiches, my salads, my p",
]

prompts = [
"Simply put, the theory of relativity states that ",
Expand All @@ -778,24 +759,22 @@ def test_compile_static_cache(self):
# Dynamic Cache
generated_ids = model.generate(**inputs, max_new_tokens=NUM_TOKENS_TO_GENERATE, do_sample=False)
dynamic_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
self.assertEqual(
EXPECTED_TEXT_COMPLETION[expected_text_completion_idx], dynamic_text
) # Both GPU architectures have the same output
self.assertEqual(EXPECTED_TEXT_COMPLETION, dynamic_text)

# Static Cache
generated_ids = model.generate(
**inputs, max_new_tokens=NUM_TOKENS_TO_GENERATE, do_sample=False, cache_implementation="static"
)
static_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
self.assertEqual(EXPECTED_TEXT_COMPLETION[self.cuda_compute_capability_major_version], static_text)
self.assertEqual(EXPECTED_TEXT_COMPLETION, static_text)

# Static Cache + compile
model.forward = torch.compile(model.forward, mode="reduce-overhead", fullgraph=True)
generated_ids = model.generate(
**inputs, max_new_tokens=NUM_TOKENS_TO_GENERATE, do_sample=False, cache_implementation="static"
)
static_compiled_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
self.assertEqual(EXPECTED_TEXT_COMPLETION[self.cuda_compute_capability_major_version], static_compiled_text)
self.assertEqual(EXPECTED_TEXT_COMPLETION, static_compiled_text)


@slow
Expand Down
Loading

0 comments on commit 47e9238

Please sign in to comment.