cleanup logging

Signed-off-by: Vladimir Mandic <mandic00@live.com>

modules/attention/router.py

@@ -120,7 +120,7 @@ def build_sparse_stage(plan: Plan):
         return None
     built = sparse_stage.make_stage(options)
     if built is not None:
-        log.info(f'Sparse attention: backend={capable[0]} budget={options.budget:.0%} gate={options.min_tokens} schedule={options.schedule_steps}x+{options.schedule_bump:.0%}')
+        log.info(f'Attention: sparse=True type={capable[0]} budget={options.budget:.0%} gate={options.min_tokens} schedule={options.schedule_steps}x+{options.schedule_bump:.0%}')
     return built
 
 

modules/attention/sparse/stage.py

@@ -7,6 +7,9 @@ from modules.attention import context
 from modules.attention.sparse import layout as layout_mod
 from modules.attention.sparse.selector import BlockSelection, SparseSpec, block_count, radial_blocks, schedule, select_blocks
 
+
+debug = os.environ.get('SD_ATTN_DEBUG', None) is not None
+
 # SD_SPARSE_PATTERN=radial replaces the content aware selection with a static band around the
 # diagonal at the same density: the control the selector has to beat, and the fallback if it does not
 pattern = os.environ.get('SD_SPARSE_PATTERN', 'adaptive').strip().lower()
@@ -123,8 +126,8 @@ def make_stage(options: StageOptions):
             from modules import shared
             hit = match_exclusion(state.model_key, getattr(shared, 'sd_model_type', None), options.exclude)
             excluded[state.model_key] = hit
-            if hit: # an enabled setting that cannot act says so rather than doing nothing quietly
-                log.info(f'Sparse attention: "{hit}" is on the exclusion list; attention stays dense')
+            if hit and debug: # an enabled setting that cannot act says so rather than doing nothing quietly
+                log.trace(f'Sparse attention: "{hit}" is on the exclusion list; attention stays dense')
         return len(hit) > 0
 
     def decline(reason: str):
@@ -142,7 +145,8 @@ def make_stage(options: StageOptions):
         if attn_mask is not None and 'masked_block' not in caps: # flex would need a mask_mod to combine the two
             if 'masked' not in notified: # an enabled setting that cannot act says so rather than doing nothing quietly
                 notified.add('masked')
-                log.info('Sparse attention: this model passes an attention mask and the serving backend cannot combine it with a block selection; attention stays dense')
+                if debug:
+                    log.trace('Sparse attention: this model passes an attention mask and the serving backend cannot combine it with a block selection; attention stays dense')
             return decline('masked')
         if query.device.type == 'cpu' or query.dim() != 4:
             return decline('unsupported tensor')
@@ -152,7 +156,8 @@ def make_stage(options: StageOptions):
         if seq_q < options.min_tokens:
             if seq_q not in inactive: # an enabled setting that cannot act says so rather than doing nothing quietly
                 inactive.add(seq_q)
-                log.info(f'Sparse attention: inactive at tokens={seq_q}, below the minimum sequence of {options.min_tokens}; attention stays dense')
+                if debug:
+                    log.trace(f'Sparse attention: inactive at tokens={seq_q}, below the minimum sequence of {options.min_tokens}; attention stays dense')
             return decline('below the minimum sequence')
         budget = budget_for_step()
         if budget >= 1.0:

modules/devices.py

@@ -531,17 +531,19 @@ def set_sdpa_params():
             else:
                 sdpa_original = torch.nn.functional.scaled_dot_product_attention
         except Exception as err:
-            log.warning(f'Attention: type="sdpa" {err}')
+            log.warning(f'Attention: type="SDPA" {err}')
         try:
+            options = {}
             torch.backends.cuda.enable_flash_sdp('Flash' in opts.sdp_options or 'Flash attention' in opts.sdp_options)
             torch.backends.cuda.enable_mem_efficient_sdp('Memory' in opts.sdp_options or 'Memory attention' in opts.sdp_options)
             torch.backends.cuda.enable_math_sdp('Math' in opts.sdp_options or 'Math attention' in opts.sdp_options)
             if hasattr(torch.backends.cuda, "allow_fp16_bf16_reduction_math_sdp"): # only valid for torch >= 2.5
+                options['math'] = 'fp16/bf16'
                 torch.backends.cuda.allow_fp16_bf16_reduction_math_sdp(True)
             torch_info.set(attention="sdpa")
-            log.debug(f'Attention: type="sdpa" kernels={opts.sdp_options}')
+            log.debug(f'Attention: type="SDPA" kernels={opts.sdp_options} options={options}')
         except Exception as err:
-            log.warning(f'Attention: type="sdpa" {err}')
+            log.warning(f'Attention: type="SDPA" {err}')
         attention.install_router([opts.cross_attention_optimization], attention.Platform(backend=backend, device=device), sdpa_original)
         if report:
             report_attention()