feat(server): add frequency/presence penalty sampling parameters

[howard0su]

Add OpenAI-compatible frequency_penalty and presence_penalty to the C++ server's sampling pipeline. Implementation is model-independent and reusable across all backends (qwen35, qwen3, gemma4, laguna).

Changes:

• Add freq_pen, pres_pen fields to SamplerCfg with needs_logit_processing() helper
• Implement penalty logic in sample_logits() (applied before top-k/softmax)
• Parse frequency_penalty, presence_penalty, repetition_penalty, rep_window from HTTP JSON
• Fix temp=0 bypass bug: penalties now apply even with greedy decoding
• Update all backends to use needs_logit_processing() instead of raw temp check
• Spec decode correctly disabled when penalties are active
• Add 19 unit tests (1364 assertions total)
• Add 14 HTTP integration tests for sampling parameters
• Add API.md documenting supported/TODO parameters with priority levels

Sampling chain: rep_penalty → freq/pres_penalty → top_k → softmax(temp) → top_p → draw

[cubic-dev-ai]

3 issues found across 10 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="dflash/scripts/test_server_integration.py">

<violation number="1" location="dflash/scripts/test_server_integration.py:232">
P2: Flaky probabilistic integration test: `test_temperature_high_produces_variance` asserts `len(results) >= 2` across only 5 generations with different seeds, but legitimate sampling behavior may still produce identical outputs for all seeds — especially for a prompt like "Name a color." where one token may dominate the distribution. This creates a flaky test that can fail on correct behavior.</violation>

<violation number="2" location="dflash/scripts/test_server_integration.py:279">
P2: `test_top_p_restrictive` asserts deterministic output without a fixed seed, which can cause flaky failures</violation>

<violation number="3" location="dflash/scripts/test_server_integration.py:343">
P2: Test assertion for frequency penalty is too weak to detect regressions</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

P2: Flaky probabilistic integration test: test_temperature_high_produces_variance asserts len(results) >= 2 across only 5 generations with different seeds, but legitimate sampling behavior may still produce identical outputs for all seeds — especially for a prompt like "Name a color." where one token may dominate the distribution. This creates a flaky test that can fail on correct behavior.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At dflash/scripts/test_server_integration.py, line 232:

<comment>Flaky probabilistic integration test: `test_temperature_high_produces_variance` asserts `len(results) >= 2` across only 5 generations with different seeds, but legitimate sampling behavior may still produce identical outputs for all seeds — especially for a prompt like "Name a color." where one token may dominate the distribution. This creates a flaky test that can fail on correct behavior.</comment>

<file context>
@@ -209,6 +209,241 @@ def test_temperature_zero_is_deterministic(self):
+        content = r.json()["choices"][0]["message"]["content"]
+        assert len(content) > 0
+
+    def test_temperature_high_produces_variance(self):
+        """High temperature with different seeds should produce different outputs."""
+        base = {
</file context>

Suggested change

	def test_temperature_high_produces_variance(self):
	"""High temperature with different seeds should produce different outputs."""
	base = {
	"model": MODEL_NAME,
	"messages": [{"role": "user", "content": "Name a color."}],
	"stream": False,
	"max_tokens": 8,
	"temperature": 1.5,
	}
	results = set()
	for seed in [1, 2, 3, 4, 5]:
	r = post_json("/v1/chat/completions", {**base, "seed": seed})
	assert r.status_code == 200
	results.add(r.json()["choices"][0]["message"]["content"].strip())
	# With high temp and different seeds, we expect at least some variance
	assert len(results) >= 2, f"Expected variance, got: {results}"
	# Removed: testing statistical variance with small sample size is inherently flaky
	# Temperature correctness is covered by test_temperature_nonzero_accepted,
	# test_temperature_zero_is_deterministic, and test_seed_reproducibility.

[cubic-dev-ai]

P2: Test assertion for frequency penalty is too weak to detect regressions

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At dflash/scripts/test_server_integration.py, line 343:

<comment>Test assertion for frequency penalty is too weak to detect regressions</comment>

<file context>
@@ -209,6 +209,241 @@ def test_temperature_zero_is_deterministic(self):
+        text_no_pen = r_no_pen.json()["choices"][0]["message"]["content"].lower()
+        text_pen = r_pen.json()["choices"][0]["message"]["content"].lower()
+        # The penalized version should have fewer repetitions of "hello"
+        assert text_no_pen.count("hello") >= text_pen.count("hello")
+
+    def test_presence_penalty_accepted(self):
</file context>

[cubic-dev-ai]

P2: test_top_p_restrictive asserts deterministic output without a fixed seed, which can cause flaky failures

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At dflash/scripts/test_server_integration.py, line 279:

<comment>`test_top_p_restrictive` asserts deterministic output without a fixed seed, which can cause flaky failures</comment>

<file context>
@@ -209,6 +209,241 @@ def test_temperature_zero_is_deterministic(self):
+        assert r.status_code == 200
+        assert len(r.json()["choices"][0]["message"]["content"]) > 0
+
+    def test_top_p_restrictive(self):
+        """Very low top_p should make output more deterministic."""
+        body = {
</file context>